GLM 5

[safe hill]

https://chat.z.ai/

[glacial onyx]

we got the open source week

[paper meadow]

https://tenor.com/view/happy-birthday-pony-gif-27170110

[glad scroll]

hi!

[worn depot]

https://glm5.com redirects to pony alpha lol

[heady shadow]

yay

[ebon depot]

its out

[worn depot]

on hf?

no ,not yet

[ebon depot]

on api

[worn depot]

oh

[placid crescent]

😮

[sinful echo]

can 'thinking' be turned off with this? still haven't seen any docs on z.ai's website

[placid crescent]

(they better make a glm-5-flash for me >v< )

[stable salmon]

Now the battle if this is better the same or worse than Pony Alpha.

[pseudo anvil]

im nervous

[ebon depot]

things are going well on the z.ai server

[paper meadow]

Most polite way to talk with laowai

[plain topaz]

only on max

[glad scroll]

https://openrouter.ai/z-ai/glm-5

[half canopy]

https://docs.z.ai/guides/llm/glm-5

Wake up babes, the docs and graphs for glm 5 are up

1 dollar input and 3.2 dollar output pricing

[ebon depot]

more expensive than kimi 2.5

[half canopy]

It's better than kimi 2.5 from what I tested, it also has more active params from what has been leaked.

[half canopy]

Considering it performs on the level of opus 4.5 thinking, it's not that expensive

[pseudo anvil]

rp mention 🔥

[rotund wing]

yey

[paper meadow]

Open weights?

[ebon depot]

presumably, other providers have it

[half canopy]

Agents, coding and gooning(RP). It covers the trinity of LLM usage 🍾

[chilly leaf]

any providers support nothink?

[worn depot]

benchmarks are here on the api and blog now

https://z.ai/blog/glm-5

though weird they dont compare to k2.5

[worn depot]

nvm they do lower on the page, just not in the main benchmark image

but its more expensive than k2.5 tho

[cobalt furnace]

No GLM-5 on the Pro coding plan 🥀

[worn depot]

we’re rolling out GLM-5 to Coding Plan users gradually.

[half canopy]

It is slowly being added

[worn depot]

max get it right now, but i guess its a slow rollout for the others

[half canopy]

And will consume more credits than glm 4.6

Also considering it's 4x the price of Kimi 2.5 in their own benchmark, doesn't look that good even if it's 2-3% better

😭

[worn depot]

i honestly never liked glm models, but this one atleast under pony alpha was quite good

[half canopy]

The models directly using the api aren't worth the money, but if you buy their coding plan. For 9 dollars for 3 months, you can use like 30 mil of something tokens every 5 hours

Super high value

[light dust]

Against kimi?

[half canopy]

On the most basic plan

[worn depot]

in general, i didn't really compare them on the same tasks
but it felt about as good

[light dust]

Unrelated but what do you use for agentic coding? Kimi or glm

[worn depot]

kimi

[ebon depot]

not with glm 5 anymore

[worn depot]

and minimax for simpler stuff

[light dust]

Thinking or 2.5?

[worn depot]

2.5

[light dust]

Thanks

I never got minimax to work for me

It really feels dumb

[kind rover]

lets see how this model does

[half canopy]

They're slowly rolling glm 5 to the coding plan, I believe

[light dust]

But if the translation could be Opus = kimi 2.5, Sonnet = Glm, Haiku = Minimax

[ebon depot]

not to lite

[half canopy]

oO 😭

[inland heath]

Ow, this is pricy

[worn depot]

i wish, but definitely not opus level, but better than sonnet

[inland heath]

Cheapest provider is $0.80 / $2.56

[light dust]

Yeah unfortunately no os model

Comes close to the one shot performance of opus

[ebon depot]

https://tenor.com/view/monkey-monkey-sleep-sleep-wrapped-up-tucked-in-gif-5999802737525363478

[light dust]

But opus is so damn expensive 😭

[light dust]

Congrats on the mod role bro

Momentum got you something nice

[ebon depot]

the lgbtq+ community has forgiven momentum

[light dust]

Glm 5 is pretty nice, at least the hidden version of the model worked nice with opencode

[inland heath]

Seems like the Z.ai server is in the middle of something

Some stuff about the pro plan changing prices

[plain topaz]

glm5 very slow rn

[muted drum]

Can't use it...

[tribal elbow]

thank you!

[cerulean bough]

day i upgraded to max I got booted, no idea why was actually defending them against some guy going crazy lmao..weird serve

[ebon depot]

https://huggingface.co/zai-org/GLM-5

[summer gulch]

It's actually a bit of an upgrade - Pony struggled to consistently add colour tags within story dialog, but this actually seems to be managing it so far.

[warm silo]

it's not just that, they also changed some other things about their plans such as adding a weekly limit and removing the "flagship model updates" promise from the pro plan, glm-5 is currently on the max plan only (which is now $80 a month)

all plans got price hikes as well, Lite was $6 a month and now is $10 for the same usage and models

[inland heath]

Oh, boy

[kind rover]

Yeah this model isn't that good for me

[sinful echo]

Can we disable thinking?

[rare hedge]

LOL.

who would pay $80 a month for a shitty chinese model instead of $100 for opus 4.6

[kind rover]

heres what I got out of it https://x.com/Gardasio/status/2021643274952618251

I think this model has potential just that compute is just not there yet

maybe if cerebras can host it without cooking it too badly it would have potential

[ebon depot]

yeah z.ai is clearly compute constrained

[kind rover]

but yeah still far away I think from being usable day to day

[kind rover]

100%

[wooden briar]

https://chat.z.ai/space/q1aa81f5fz31-art

Awesome

(ghost in the shell reference)

[merry plover]

lul, this model tax is insane

(it's basically the same model architecture!)

[mellow ember]

LOL IPO

[drowsy kettle]

How is it for roleplaying?

[ebon depot]

very good

[mellow ember]

I'm liking it so far

[velvet patio]

Pony was good so assuming it didn't get neutered probably solid

[merry plover]

claude thinks glm moe should be only ~10% more expensive : )

[sinful echo]

Can we disable reasoning here?

[velvet patio]

Bulbasaur just offered to host GLM 5 on OpenRouter for 10% more cost than GLM 4.7. What a legend

[merry plover]

GLM 5 performs 4 points higher than Claude Opus 4.5 in TerminalBench

[merry plover]

if there was an easy way for me to just pay a lump sum -> rent out to openrouter, I would

tokens need to have property managers

[cerulean bough]

oof those new sub prices.

[cobalt furnace]

I see that lite has gone up, but I'm pretty sure Pro stayed the same at 30USD/mo

[chilly leaf]

5 at least writes a lot better than 4.X, much less rigid and ism prone

was a little worried it would go the way of deepseek

[placid crescent]

what happened to deepseek?

[chilly leaf]

versions after 0528 became much more sterile imo

[cerulean bough]

huge upgrade on 4.7

can use 3x concurrency on 4.7 now with max plan

but i want the shiny new toy

[half geode]

The coding plan stuff is going to burn some goodwill

[cyan sequoia]

its a much better model though

so makes sense for them to charge on how good it is while still overcutting the ones that do match / out perform it by a good margin

[half geode]

I get it, the model is bigger and compute is constrained, price hiking is understandable, but not giving access to the people who already paid is what will piss people off

[cyan sequoia]

This feels like sonnet 4.5

but 1/6 ish the price

maybe even better than sonnet

[faint sphinx]

Hello everyone

[cerulean bough]

huh? I have max plan until the end date..did they say something about the change?

[cerulean bough]

gm

[half geode]

What? It is 3x cheaper, not 6x.

[half geode]

Max has access rn. Pro supposedly will in the future. Lite will not.

[cerulean bough]

Oh i get you, lite and pro dont have 5 access, thanks for clearing that up

[cyan sequoia]

$1 $3.20 $0.20 vs $3 $15 $0.3

that is input / output / cache

total its a bit more than 5x cheaper

[fluid zealot]

You'll have to adjust for verbosity. Cannot guess a cars mileage (tok/task) while only accounting for gas type (mtok).

[half geode]

Ah, I missed the output price, though it was $1/$5 not $1/$3.20. My B

They're potentially trying to push to Lite I guess? The tweet implies that while the website does not.

[wooden briar]

And Claude's verbosity is HIGHLY variable based on the output effort params, at least with the new Opus

[half geode]

GLM-5 is coming to Coding Plan Pro users within one week, and we're working to bring it to everyone after that.

[graceful tapir]

no arc agi?

[signal oar]

can anyone who has used it answer if glm 5 is fire

[graceful tapir]

no ofc its not lmao. look at the gpus they are using

[signal oar]

They using gtx 1070 ti for ts?

[graceful tapir]

yes. that is trash for llms

[signal oar]

y'caint blame em i mean they cant get them fire gpus in china

[graceful tapir]

glm is not hot shit. its just shit.

because trash gpu.

[signal oar]

give me a bunch of gpus and i'd come out with some bullshit though 😂

[thorny moss]

That is one hell of a brainlet answer lol

GLM 5 is absolutely solid, so far agentic and coding workflows have been reliable and coding is nothing it needs to hide for in front of GPT 5.3 Codex and Opus 4.6

Given its price, it's probably going to be a great general / default model, with specific domains maybe delegated to other models. Outside of productivity, it's also quite the pleasant RP model.

For now I would wait for more reliable providers to get available, so you have some buffer in case one isn't doing well

[fluid zealot]

unlike 4.7, no chess reasoning loops 🥳

[thorny moss]

If rumors are true, they mainly use Huawei chips. While not as powerful as Nvidia, being able to train and serve at all on such hardware is an achievement. But that's based on rumors.

[thorny moss]

Overall reasoning feels a lot less overbearing, whatever they changed is welcomed xD

[fresh osprey]

Any changes between pony and 5?

[paper meadow]

Not many, they just horsing around

[thorny moss]

At least regarding possible censorship additions, no. At least in my test suite it didn't complain (which is good)

[chilly leaf]

man some of the third party providers are getting hammered

[mossy charm]

Its quite good IMO

Lite doesn't include 5 yet tho 🙁

[wet comet]

nice

[wet comet]

we love ur models

are yall releasing a small model soon?

[pulsar niche]

Underwhelming 56.2% on lateralbench, underperforming K2.5

[wet comet]

just realised this is a good opportunity to mention that i was permabanned in your server on accident

please resolve this

[mossy charm]

Try it, it's a pretty noticeable upgrade from 4.7 IMO

[chilly leaf]

I've been trying it and liking it, the issue is that my preferred provider is getting swamped right now lol

[half geode]

Interested to see how this plays out at the pricing

It costs roughly double what Kimi 2.5 and Gem Flash 3 do, and those models are no slouches.

[wooden briar]

.2% diff is within statistical error range, hardly a difference. What's that bench anyway? I googled it and all the results were stuff about workouts, lol

Edit: I was looking at the wrong Kimi model on the chart, error on my part, disregard

[inland heath]

https://www.lateralbench.org/, weird that it doesn't have good SEO. Really cool benchmark

[proud river]

Is this available on OR?

[ebon depot]

yes

[wooden briar]

Thanks, looks interesting. I do like 'weird benchmarks', like the one that tests models on how well they do on the NYT Connections puzzle

[proud river]

Nice, good model. And the old jailbreak still works

[pulsar niche]

Thanks. It was a weekend project and turned out better than expected so I keep running new models. I'm a backend guy, not a frontend guy. If you do have suggestions I'm open to them

[mossy charm]

It really is pretty good

Its better than Gemini Flash imo

Closer to Sonnet

(On major languages)

[tall wasp]

is artificial analysis any good of a benchmark? GLM 5 is scoring above kimi k2.5, gpt 5.2 codex and gemini 3 pro

[merry plover]

it's ok but it's very biased

[ebon depot]

it's gotten better, used to be way worse

[merry plover]

yeah

but it's clearly still not a very "general" benchmark

for example, it doesn't very well show you that GLM 5 with hallucinations is still complete dogshit

compared to any frontier model

[ebon depot]

generally my favorite benchmark is https://fiction.live/stories/Fiction-liveBench-Feb-21-2025/oQdzQvKHw8JyXbN87

[fresh osprey]

I keep getting routed to Friendli, which take the input, then immediately stops..

[merry plover]

I think one of the best "aggregate" benchmarks might be "what is the average normalized performance of the model on ANY benchmark you give it"

because its generalized performance is one of the best indicators for its performance in "real usage"

so yeah, picking random benches like fiction live bench are a good proxy

especially if they don't seem to be getting bench maxxed

[ebon depot]

true, I just like fiction livebench because it tests narrative understanding instead of needle in a haystack

[tall wasp]

i see, makes sense, i still find kimi k2.5 to be the best open model currently

[tall wasp]

from my testing zai themselves seem to be the only reliable provider rn

[merry plover]

I personally find any model with awful hallucinations really really hard to use, so yeah I agree, only because kimi k2.5 is "ok" on hallucinations

[ebon depot]

official providers are almost always worth it

[balmy cave]

speed is atrocious

[halcyon estuary]

TTFT is like 30 seconds

on venice

[thorny moss]

Huh, I'm getting around 20tps on Z.AI and Novita

Give it some time, until everything is stable it usually takes a bit. At least Anthropic and OpenAI have been equally as slow and unreliable this week...

[halcyon estuary]

looks fine now

[jagged kayak]

is there a way to set "clear_thinking": False? i tried extrabody but not workinghttps://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking

[fluid zealot]

unfortunately every model release nowadays is always crippled by slow inference. annoying indeed. also makes benchmarking endeavors very tedious. everyone is hardware-starved https://x.com/Zai_org/status/2021656633320018365

Z.ai (@Zai_org)

GLM-5 is coming to Coding Plan Pro users within one week, and we're working to bring it to everyone after that.

To be upfront: compute is very tight. Even before the GLM-5 launch, we were pushing every chip to its limit just to serve inference. We appreciate your understanding

X · Feb 11, 2026, 6:44 PM

[lone halo]

codex got faster in the last 5.3 release as Stargate Project datacenters are coming online

[thorny moss]

It also was down for half a day, at least in my region xD

[pseudo anvil]

still toying with GLM 5 but I find myself preferring pony alpha

probably placebo or bias

[ebon depot]

interesting, I'm the opposite

I like glm 5 release better

[pseudo anvil]

one thing I noticed is that it’s very context dependent

I always find it gravitating towards the style of the intro msg as opposed to my writing style prompt

[ebon depot]

probably a prompt difference then, I send the chat history as a user message instead of actual multi turn

[pseudo anvil]

ah interesting

[cloud stag]

Not that soon

[half matrix]

how is it?

compare to the previous checkpoint

pony

[cyan sequoia]

the same

[tall wasp]

hallucination problems aside, how good is glm 5 on web dev? have anyone tested that?

[safe hill]

https://modal.com/glm-5-endpoint

[merry plover]

How about faster inference? :D

[ebon depot]

how are they gonna do that in china 😭

[ancient flare]

i did rp with 4.7, then with 5, it was night and day

[ancient flare]

howahway chips!!

[cloud stag]

It's getting faster as it's a new base model. There are lots of optimizations we need to do.

[wet comet]

glm 4.7 flash is still the best size to performance coding model IMO

[unkempt hemlock]

Thank you very much for your recognition～

[paper meadow]

It's clever model, but it repeats itself after retries, as if Temperature is set too low, making answers deterministic. But I checked, temp is working right, properly breaking after 1.2+

Only me?

[arctic portal]

No glm is always like this

every other model has been like this

try changing samplers

[paper meadow]

I didn't try other GLMs thoroughly

[paper meadow]

There are not many samplers in API

[arctic portal]

glm follows instructions very rigidly

perhaps try not using a system prompt

[paper meadow]

Oh, maybe I need introduce some randomness in instructions or make statements more vague

Good idea

[paper meadow]

Huh, it means I need separate system prompt, completely rewritten, just for GLM5, while others work from good to okay with current one. Tough choice

[fluid zealot]

Tested GLM-5:

Beefier GLM hybrid-reasoning MoE model (355B-A32B → 744B-A40B).

Default/Thinking:
Slightly more verbose than previous GLM models, DeepSeek-R1 0528 level.
76% of tokens were generated during reasoning.

• very high general logic and reasoning
• I saw no leaps in my STEM & tech tasks
• reasonably censored
• Unlike 4.7, no reasoning loops encountered

Chess performance wasn't great in a vacuum: 6k tok/move ~780 mixed elo /w 62% accuracy, decent blind legality, around o1-mini. Best among GLM family though.

Nonthinking:

• ~76% token savings (non-reasoning segments were samey).
• negatively impacts logic and maths
• was slightly less likely to refuse in censorship testing

Overall, very solid and one of the best open models currently, but YMMV.

[paper meadow]

For me, thinking verbosity went from K2-like (HUGE) to compressed more average one, during the Pony Alpha phase, like as if they updated model with less verbose reasoning version

[fluid zealot]

how verbose a model is depends on the task. this is just the average from the 250 general use queries. if I isolate to chess moves, kimi is ~50% more verbose in one mode (full info), but ~50% less verbose in another (no info). So obviously one should compare it for their own use case outcomes.

[burnt belfry]

no need

There's always gonna be outliers when your tasks are too specific and not diverse enough, but there's no need to highlight them. What matters is the overall difference in model behavior rather than your oddly specific tasks favoring certain models

[worn yarrow]

hooray they fixed the loops

[glacial onyx]

yeah there is literally no model that comes close for a 30b model

[pseudo anvil]

which provider do you guys use?

[ebon depot]

z.ai

[pseudo anvil]

ive been on zai too so far but considering making the switch to fireworks

[ebon depot]

why?

[pseudo anvil]

performance wise is supposedly the best from OR stats

[thorny hinge]

I had a poor first experience with GLM 5 and I think I realized why. I used it for multilingual writing, but it tests signficantly worse for this task than GLM 4.7, GLM 4.5 on NCBench, mirroring my impression that even made it feel like a ~70-120B model at times. https://www.nc-bench.com/tests/language-writing

Language comprehension is retained though, no regression there, which is interesting.

[pseudo anvil]

I am so bipolar about this model sometimes the writing is just chefs kiss then other times it has no substance

[midnight crow]

GLM 5

[ebon depot]

💔

[cobalt bronze]

poor plunger is a little slow

[paper meadow]

0.6 oh no

[ebon depot]

plunger is my stupid chud failson but i love him

[fresh osprey]

Well tps from Z.ai hit new lows.

[primal thorn]

Getting a lot of 429s

[fresh osprey]

Suddenly getting failed to stream errors

[mossy charm]

GLM 5 is only good in English or Chinese mostly

theyre getting slammed

[paper meadow]

Is it? I thought the data for training is so sparse now, they fit whatever language they find into model

And semantic vectors in models are, like, very universal

[unkempt hemlock]

We will handle it promptly～

[velvet patio]

Doing god's work

[worn yarrow]

it's a little underwhelming when you consider that they went from 355B to 744B parameters for this amount of improvement

also deepseek has really got to catch up. they are way behind considering their model is 685B params

[half matrix]

so i has done some testing with this model

at the current time with not that complex code base it able to perform really well, in term of cost to performance. it's actually pretty fair when being compare to other model with similiar capabilities.

but i hope the infra from the providers could improve so it could be cheaper, i mean i know model with more active parameters and cost less than this so yeah.

[merry plover]

GMI Cloud seems to actually be serving this model at quite decent speeds

[worn depot]

do their tool calls work though?

[burnt belfry]

DeepSeek nailed their model size honestly. GLM needing to up the size only shows that they got it wrong the first time. And 744B vs 685B is not really something you would notice, those are equivalent in practice.

Downsizing is normal since it follows the general progress, upsizing is not however. You would mostly only do that if the earlier arch was flawed.

With 700B you can just about have SOTA model in 2026. But in 2023 you would have absolutely needed 1T+, and MUCH more activated params. Much more expensive inference per token too.

Kinda just goes to show how far ahead of everyone else GPT4 was at the time tbh

[velvet patio]

Still wild to me that GPT-4 was such an important part of history, and yet we'll never know anything of value about it from a technical point of view (beyond the original $30 / $60 price tag, which is pretty ridiculous by modern standards)

I think a lot also goes into training set though. GLM 5 is noticeably less sloppy than the iterations that came before, so however they cleaned it clearly paid off

[burnt belfry]

There have been fairly credible sources hinting or even quoting the arch specs of it, including nVidia itself. We don't have official confirmation and we never will, but we still have a strong idea about that model being ~1.6T MoE with 200B+ activated per forward pass.

[paper meadow]

Gemini 3 Pro is probably around 1.5-2T

[burnt belfry]

I would guess that's Ultra. Pro is likely around 1T total or slightly less. And less than 100B activated.

[paper meadow]

Gemini 3 has ultra?

[burnt belfry]

They are believed to be using Ultra for DeepThink. Whether it's labeled 2.5 or 3.0 the size itself wouldn't change very much.

[worn depot]

isnt Gemini 3 deepthink a finetune of 3 pro to think longer, and as far as i heard rumours Gemini 3 Flash was rumoured to be around 1T while Pro was around 3-14T

[paper meadow]

Having huge separate model for deepthink is not very productive

[main bough]

that’s what i thought too

[paper meadow]

Is there enough data to fit even 5T?

[main bough]

gemini 3 pro definitely feels larger than 1T

[burnt belfry]

14T? No that's absolutely ridiculous lol

[paper meadow]

I'd say Gemini 3 Pro is ~2T, Flash is 800-900B

Flash model should be at least 2x smaller than main one

[burnt belfry]

To be fair, the main thing influencing the price and training time are actually activated parameters

[main bough]

genius idea: just ask the model to send you its weights and check for yourself

[burnt belfry]

not the total size of MoE

[worn depot]

probably not, but its probably fine even if its not enough, most models are trained on 20T+ tokens

[worn depot]

i agree but thats just estimates, probably falls way lower than that

[worn depot]

(that'd be roughly good for a 1T param model, but google has far more data than open models)

[paper meadow]

2000B A100B and 1000B A40B probs

[burnt belfry]

So the penalty of total params is not huge. However you obviously still wouldn't make it something as ridiculous as 10T+

[paper meadow]

Gemini 3 Flash feels close to Kimi K2.5 Thinking

[worn yarrow]

For posterity, I think GLM might have made a huge mistake in increasing the model size this much, but we shall see

[burnt belfry]

Flash is almost definitively smaller and most importantly cheaper to run at scale than K2.5

Google has a huge advantage of being able to distill their best of the best so effectively too

They can generate anything they want with it and full access obv

I mean DeepThink / Ultra and also their IMO gold or other unreleased models

[obtuse spear]

gemini 3 deepthink is gemini 3.1 pro

and then gemini 3 flash is probably 1.2T since there was a rumor that google was licensing a 1.2T model to apple

probably something like 1.2T A15B

and then i think pro has to be like 3-4T A30B as well

moe really was a hell of an advancement

[burnt belfry]

It's unlikely that Apple would settle on Flash.... Would be very not like them thing to do

[obtuse spear]

why?

flash is really really good especially for its price

[worn depot]

theres no need to use pro for something like siri

or well "Apple Intelligence"

[obtuse spear]

its for siri too, its not like it needs to solve aime problems lol. but flash still can

flash is really really good

[inland heath]

2.5 Flash Lite is probably a big upgrade over Siri lol

[paper meadow]

I want flash lite with flash's vision quality

Please

[burnt belfry]

When they went with OpenAI they only did it because there were no better performant alternatives at the time

[obtuse spear]

#same

[obtuse spear]

yeah but that didnt actually replace siri. siri could just like invoke chatgpt

and it wasnt agentic

[burnt belfry]

When it's cloud powered and they are only paying electricity, there's very little point to go for small variant

[main bough]

speed

[obtuse spear]

yeah

also "small" is relative at 1.2T lol

[worn depot]

yeah but electrcity is variable based on how many peope are using it, and to lower electricity (maximize efficiency) you need high batch size to run the model on as little gpus as possible

[obtuse spear]

flash does great speed. and for siri you want people to like get the result of their task asap

makes for a better UX

[paper meadow]

Well Apple users should learn to be more patient

[burnt belfry]

They want for it to beat competition. Not merely work. I don't believe choosing Flash would work there tbh

[obtuse spear]

gulp

[worn depot]

yes but flash id say is overkill for what people ask siri, and even if it fell back to 3 pro for harder questions i think it would be good

they allegedly trained a model as big as 100B~ params internally but it wasnt good enough

[obtuse spear]

i actually think that gpt 5 mini is overlooked

[burnt belfry]

Fallback to Pro would make sense in theory, except they were fairly explicit on a deal for singular model iirc

[obtuse spear]

and yeah flash is good enough for that

[burnt belfry]

Maybe it's instead they gonna still rely on their in-house models for easy questions

[obtuse spear]

throw on some medium thinking level and youre good with 99% of siri queries

[arctic portal]

what do you use it for? i cant think of any use case where other models wouldnt be better

[worn depot]

its wayy too slow for a mini atleast via openai api its at like 20-30 t/s, and thinks a bit too long for it to be realtime like siri

gemini 3 flash is significantly better without needing to think

[burnt belfry]

And then do function calling to Gemini for more involved tasks or smth

[arctic portal]

something like gpt oss would be good for siri

[main bough]

96 tps with openai api

https://openrouter.ai/openai/gpt-5-mini

[obtuse spear]

i do classification, extraction of documents, some agentic deep research etc

[worn depot]

ok this is recent, it was never this fast.

[burnt belfry]

The thing that killed them with OpenAI was latency

Not gonna have this problem when they are hosting Gemini themselves

[arctic portal]

why wouldnt you use deepseek

[worn depot]

last time i was using it, it was about as slow as gpt 5

[worn yarrow]

Can we move this chat to the other areas

[obtuse spear]

100% agree. i dont recommend 5 mini for anything realtime/agentic/etc but my tasks are more like slow/can take a while. just need to get the right answer

[obtuse spear]

vision

[arctic portal]

qwen vl has very good vision

[main bough]

moving chats when they’re already active is cumbersome and kills the flow, personally I think we should embrace conversations naturally arising in unintended places 😊 👍

[paper meadow]

Gemini 3 Pro -> 3 Flash -> Qwen3 VL 235 instruct -> GLM 4.6V for vision tasks

[burnt belfry]

Kinda crazy how much naming affects things. The current lineup is no different to o3 vs o4-mini

[paper meadow]

Maybe Kimi K2.5 is around 4.6V or better for vision

[burnt belfry]

they were both same gen

[obtuse spear]

where does k2.5 land in here?

oh okay

i havent tested it at all

[arctic portal]

kimi has way better vision than 4.6v

worse than gemini 3 pro

[paper meadow]

With or without web search?

[arctic portal]

without

[paper meadow]

I wanted to check K2.5 vision in OR, but forgor

[obtuse spear]

a little too retarded for my usecase

[paper meadow]

On site it did good, but maybe it was with external tools

[burnt belfry]

And now people overlook gpt5-mini, because the naming already strongly implies worse performance

[obtuse spear]

it needs to extract -> transform with some specific specs etc

[worn depot]

chatgpt still falls back to it when you run out of GPT 5.2 messages and have thinking enabled and its pretty good

[paper meadow]

https://dubesor.de/visionbench

[obtuse spear]

yeah with thinking like level=high gpt 5 mini is really really solid

[paper meadow]

GPT5 is unusable for its censoring, otherwise I agree with dubesor

[obtuse spear]

unfortunately its a bit slow at effort=high but even medium is solid

[burnt belfry]

No surprises for it being good. When you look at the numbers the difference is near exactly the same as for my mentioned o3 vs o4-mini

[obtuse spear]

idk abt this one. at least for me gpt 5.2 has really really solid vision

[paper meadow]

Qwen3-VL-235B-A22B-Instruct has providers doing it 1/3 price of Gemini 3 Flash, meaning it's unbeatable for mass vision tasks where subtext and world/media knowledge is withing common limits

[arctic portal]

gpt always hallucinates in my tasks

[arctic portal]

i completely agree on this

[burnt belfry]

dubesor tends to do overly specific tests that favor his very specific workflows. Falls near anecdotal findings

[pseudo anvil]

you're absolutely right!

[arctic portal]

youve hit the nail on the head

[obtuse spear]

i agree but if you pay a bit more then ytou go to gpt 5 mini

[paper meadow]

Dubesor's vision test are technical, as far as I understand, so it does not require world knowledge for model, so 8B with good vision can do good even if base model is stupid

[paper meadow]

Gpt is worse on vision compared to 3 flash

Not mentioning it will cry about naked shoulder

[burnt belfry]

Well they do have objectively and measurably much less hallucination than Gemini3 so there's that. But obviously you can't eliminate it completely.

[obtuse spear]

its easy for us to say that gpt 5 mini is super underrated because the number of tokens that OR processes for it is low, but we need to remember all of the people who use openai api directly is insanely huge

[arctic portal]

3 flash is 50% more expensive than mini

[obtuse spear]

yeah, but more expensive

[obtuse spear]

what is naked shoulder

[paper meadow]

Erotica

[obtuse spear]

okaaaaaaaaaayyyyyyyy

i think we have different use cases

[paper meadow]

I meant, it overreacts about not just suggestive images, but anything that can be counted as not for kids

[obtuse spear]

and im over here just processing documents 🥀

what a primal use of ai

[paper meadow]

Just use OCR than. Or you will send something like Book of Vile Darkness or Monster Manual and some overaligned vision model tells you are criminal for forcing vision to read into it

[obtuse spear]

cant ocr diagrams

[arctic portal]

gemini 3 flash might be less expensive than 5 mini

if you dont use reasoning

[paper meadow]

Don't forget caching

[arctic portal]

but then i dont know how good it is

[obtuse spear]

google caching 🥀

[paper meadow]

Reasoning for vision is not useful and sometimes hurting

[inland heath]

I remember the old days when it was easier to measure a LLM's cost

No reasoning, no caching

[obtuse spear]

caching has been around for a while

at least with openai

but now you have to consider how good a providers caching is

[halcyon estuary]

old days i only used openai or anthropic

[obtuse spear]

and when it comes to google 🥀

whenevr i use gemini i genuinely dont even take caching into consideration i just think of the input price

[arctic portal]

i dunno if you go way back to llama 3 then it was basically on par with closed source

[obtuse spear]

do ygs remember the llama 3 leaked weights

[paper meadow]

https://tenor.com/view/oh-i-member-south-park-s20e1-member-berries-member-gif-20603952

[burnt belfry]

Gemini doing not good at all

GLM5 actually impressive

[arctic portal]

gemini has the most knowledge though

it just always makes up bullshit if it doesnt know

[obtuse spear]

yep

the knowledge is insane

so impressive for "where was this photo taken" quiestions

[burnt belfry]

They are trying to force grounding to "fix" it which is crazy

Every single model will hallucinate less with search

that's not a fix

[arctic portal]

it just goes insane over the date being 2026 and will start rambling about fictional timelines

[obtuse spear]

google's "google search grounding" is top tier shit

[burnt belfry]

I mean like I said... if you enable search for OpenAI or nearly every other lab, it will improve drastically over baseline as well

[obtuse spear]

oh yeah im not defending them at all

[arctic portal]

kimi has the best search and deepresearch imo

[obtuse spear]

but i mean compare google search grounding vs providing your own tools vs openai native search

[arctic portal]

they have some custom crawler

[obtuse spear]

good night google

kimi?

[arctic portal]

yeah its really good at search

[obtuse spear]

please elaborate because im building a deep research agent rn

[arctic portal]

and it does tons of searches

well im not using it for my own tooling, i just use it off the site

[obtuse spear]

you can probably use it through OR though right?

[arctic portal]

nope only through their site

[obtuse spear]

see this is what really pisses me off about gemini. you give it tools liek web_search and web_fetch and then it just makes three web_search calls and decides its done

i can never depend on it doing really good research

[arctic portal]

well actually nevermind moonshot provides search in its own api

i dont know if you can use this over OR

[obtuse spear]

yay

okay lets try

[arctic portal]

their pricing is better than OR, only 0.005$ per search

[obtuse spear]

nope you cant

[burnt belfry]

it works great for AI Mode and Overview where you need speed, but for Gemini itself I'm not quite convinced

[obtuse spear]

not on OR

[burnt belfry]

Seems like it's less in-depth than chatgpt

[obtuse spear]

its terrible

[arctic portal]

do you have tons of money dumped in OR? just register on their platform

[obtuse spear]

if i really need something in depth i know that gpt 5.2 xhigh got me

it does TONS of searchjes

[obtuse spear]

not really but i use crypto payments

does moonshot support that?

[arctic portal]

prolly not

[obtuse spear]

also those 10M free gpt 5 mini tokens are a godsend

[worn yarrow]

just an FYI to those on the coding plan, I didn't get an announcement about this and only noticed just now
GLM-5 uses 2-3x quota. Can't see anywhere that the off-peak v. on-peak is published

[obtuse spear]

kinda makes sense but sucks that they didnt say

[umbral stag]

@unkempt hemlock @faint sphinx

For the future..

I advise the team if they want to make model that tight in moderation then do it in a way that didn't compromise the freedom of the people.

One of the way is by improving it's instruction following capability, then the team can using system prompt that also injected to the API to ensure moderation for the official endpoints while allowing the other endpoints that come from non-official to have more freedom.

It will allow the team to have better legal power while giving the people what they want, freedom from any corporate morality.

xAI doing it with their grok model, they didn't have model that strict with moderation but they making it strict at following instruction, that why grok show distinct behavior either from one platform to other, but they always inject the moderation system prompt in all of their official products include the API.

At the end of the day, it's the team right to decide.
I just giving my piece of mind, always hope the best come to Z.ai team.

Thanks for reading

[halcyon estuary]

what are you being restricted?

[pseudo anvil]

i havent had any censorship issues myself

[umbral stag]

I am talking about the future, i enjoy using GLM model as agents to scower through the internet and i support Zai team venture in this industry.

We need to reminded our mind about their previous model GLM 4.7 thinking traces and their path to be public company.

Also don't forget about the tightening of the laws too.

There are sings for the future,

---

GLM 5 is better at being uncensored than any previous generation of GLM but still it didn't gonna stop the tightening of the industry.

I think, we as human, after we reading something we need to spend few minutes thinking about what we just read, finding more information and insight that are deeper than the surface.

[halcyon estuary]

ok valid

even anthropic has that fear

of being too restrictive with our friendly neighbour Claude

read the latest constitution

and honestly

they're pussies

[unkempt hemlock]

Thank you for your patience. Our official team will continue improving and working hard to resolve the issues for everyone~

[mossy charm]

it was similar to that w/4.6 & 4.7 etc

Where the newer one was more

[worn yarrow]

i dont remember that happening for 4.7

[ancient flare]

glm 5 has amazing vibes after trying it out btw.

[pseudo anvil]

honestly the dialogue is some of the best I’ve seen never have i been so engaged

whatever yall did to bring characters to life keep cooking

[ancient flare]

truly, it really is "diet claude" and the best for open weight roleplay. glm 5 is so cozy.

[worn yarrow]

GLM-5 & https://github.com/nich2533/just_say_no

[paper meadow]

Glm5 scored very high on fiction livebench, it's almost too hard to believe the difference, Kimi K2.5 levels of context retention

[light dust]

Benchtraining perhaps?

Or do you notice it when using it

[paper meadow]

I didn't check myself with my time-travelling mind-hopping story setup

I need probably to invent something more heavy

[arctic portal]

benchtraining is impossible

look at the benching process of this test

its pretty saturated anyways

[light dust]

Spell strawberry and count the R’s 😂

[paper meadow]

Woooow, not THAT heavy

[ornate root]

Do we know when this transition ends? I'm on the coding plan and I don't have access?!

[worn yarrow]

It's pro and max plans only at the moment

[wet comet]

opussy 4.6 wher ☹️

[ornate root]

How do you find GLM-5 for agentic coding now that you've had it for a few days?

[half matrix]

Most likely being train with really good data for specific use case like creative writing

but it will be different with use case that didn't have the long context training data before hand

[light dust]

True

[umbral stag]

This model enjoy spiting out tokens
When being face with complex problem (depend on the model perspective), it's capable to consume about 3$ for just that one problem.

[light dust]

Is it akin to high reasoning effort from gpt 5.2?

[burnt belfry]

How much are we talking about? Both Opus 4.6 and 5.2-xhigh can output 128k and then fail to arrive at the answer before exhausting it so I hope it's not that lol

[umbral stag]

I use it in coding with zed, it could continue automatically but i am not sure how much tokens it produce.
At the end when it done solving the code problem it cost me about that much.

Test also with GPT-5.2-Codex and because it able to tackle it easier it consume a bit less than that

This is quite interesting for me, so a expensive model could be cheaper if it able to solve it with fewer tokens, compare to cheaper model but required higher tokens count.

[ornate root]

That was Anthropic's defense of Opus 4.6; alleged token efficiency.

[umbral stag]

In my experiences opus 4.6 enjoy yapping more than GPT-5.2-Codex

[tall wasp]

what is better, this or m2.5?

[lone halo]

yes

[naive burrow]

what is reasoning:high for GLM-5? isn't it described as simple on/off feature in their API docs?

[proud river]

I’m liking m2.5

Very good performance for the price

Especially if you hit cache

[umbral stag]

GLM-5

Kilo code CLI vs Claude Code CLI

Which one you guys chosse and why

[obtuse spear]

i havent tried kilo code cli but claude code gaps everything like by a lot so

id imagine kilo also goes under

unfortuntaely claude code doesnt work too well with OR

[ancient flare]

the more i use the more it gets better. the best open weight model, amen 🙏

[obtuse spear]

more than k2.5?

[ancient flare]

yes, i really love glm 5's overall vibes (the added censorship however is frustrating). as assistant, k2.5 thinking is very interesting model. i love debating with kimi models because they never sugarcoat stuff.

but "overall", i like glm 5 more.

and oh, kimi's coding always lacks behind despite what benchmarks says

[paper meadow]

I'd say Kimi K2.5 is better overall, but GLM5 is kinda close

[obtuse spear]

kimi coding is meh

[pseudo anvil]

censorship where

[pseudo anvil]

prose wise but glm has some of the best characterisation ive seen

[obtuse spear]

is it sycophantic

[ancient flare]

mostly in assistant.

[pseudo anvil]

ah fair

[half matrix]

ask it to make drugs

jkjkjkjk

[pseudo anvil]

tell it to be a blunt son of a bitch and u get what u ask for doesnt give u the oai glaze

[obtuse spear]

oai glaze is gone and has been since gpt 5 and arguably o3

[half matrix]

It's actually pretty cool how creative writing training actually improve model creativity at other things

compare to other models that i use to make design for site, this model win against them, the other models maybe have better coding capability but it desing aren't that good.

[light dust]

Any ideas on how to use claude code with openrouter glm5?

I’m currently using ccr but maybe there’s a better way?

[worn depot]

https://openrouter.ai/docs/guides/guides/claude-code-integration

[light dust]

They removed the section for changing models

Already looked

[worn depot]

hmm

here's an archive that still has that
https://web.archive.org/web/20260109042809/https://openrouter.ai/docs/guides/guides/claude-code-integration

[light dust]

That doesn’t work; it doesn’t respect the slug and simply uses 4.5 sonnet/opus/haiku

[worn depot]

no clue then

[light dust]

Also wondering about best practices with claude code on windows

I guess one is to install bash

[cerulean bough]

Z-code has a Claude code fork, not sure how legit it is but uses all the plugins/mcps and I think you can add OR API key

[light dust]

Thanks 🙏

[half geode]

Tbf to Kimi, it is nearly half the cost of GLM

It has done well for me in coding, but I guess I'll have to try GLM.

[light dust]

Kimi 2.5 is to glm 5????

[half geode]

Yeah? At least on input tokens. Kimi is like $0.5/$2.25 and GLM is $1/$3.25

[light dust]

Damn and usually the feeling is that kimi is better

[burnt belfry]

It's anything BUT that. When you max it out it's the most verbose (in reasoning) model that I tested from by far. #attachments message

In fact it's more verbose than anything OpenAI before 5.2 as well, in my experience.

[half geode]

Yeah. I've been a big Kimi fan since K2. It reminds me the most of Claude. Not by training on their outputs like GLM-5, but just on a core level.

[burnt belfry]

Disclaimed that I really test their limits with LLMs, but the fact is that Opus4.6 is gonna output much more than vast majority of other models once you give it a hard task it is genuinely challenged by

[mossy charm]

just use opencode

[lone halo]

how did Z-code fork Claude code when it's closed source?

[half matrix]

Is it feel like it have some awarness

[cerulean bough]

same opinion but not what OP wanted - GLM is a beast inside opencode

[light dust]

If only we had glm 5 with cerebras

[mossy charm]

I think they only have up to 4.7 right?

[light dust]

They removed it because it was only in preview

I think one of the only permanents they have is gpt-oss-120b

[mossy charm]

theyre up to 5 how is it only in preview?

[light dust]

As in cerebras was testing it out on their chips for a limited time

[mossy charm]

zai removed their general and zai chat channels lol

[obtuse spear]

they need compute for opoenai

what a joke

[royal jetty]

Anyone else having issues with Z.AI provider suddently not supporting caching since yesterday? Pretty annoying.

[mossy charm]

z.ai is basically a 💩 show lately. Their stock is through the roof and I think they have way way more people on GLM5 than they thought. They need compute desperately

Their communication on their discord has also been bad

[royal jetty]

sigh

I mean caching isn't even that hard (and saves them compute!), c'mon.

Their servers are always on fire.

[mossy charm]

I have to take a certification exam for work but I have a z.ai chat client I wrote Ill put up if I pass that

We have a ^&((** blizzard supposed to be coming through too sigh

Its not caching prompts and answers right?

[royal jetty]

Well, not caching prompts.

Don't think output caching is a thing on OR.

[mossy charm]

I created multiple themes. This is the one I use. It's retro/cyberpunk. I also have a pastels and some others.

I renamed the client too. I want to add optional calls to OR too and some others but havent gotten that far. It's mostly trivial, I'm familiar with the OR api. They all basically copied the openAI one

Its also a markdown reader because I like markdown

[mental belfry]

@peak sedge Please can you add baseten as a provider for this? they have the fastest api, the rest are slow

[worn yarrow]

They are not a good provider. I can't find the github repo where the benchmarks are but like atlascloud, they sacrifice quality outputs and calls.

[mental belfry]

is it because of the fp4?

[mental belfry]

https://github.com/MoonshotAI/K2-Vendor-Verifier

[worn yarrow]

I would be concerned they are doing speculative decoding as well somehow

[half geode]

This model is goated

Shame it's $1 mTok for inputs, but damn. Claude at home IMO.

Only weaknesses being not top-tier world knowledge, and no image input

[half geode]

But as a free WebUI model for normies I think I'd have to recommend it over the free offerings from OAI or Google or Grok which is kind of interesting.

[paper meadow]

It suffers from repetition for some reason, it's either training or attention which picks top_k options

[half geode]

In WebUI?

Not really a problem for normies regardless, I care a lot more that it doesn't hallucinate on them as hard as Gem Flash. Or get auto-routed into retardation like GPT.

[paper meadow]

Everywhere. It hard locks on certain subjects and ideas that stay the same after retrying. Deepseek v3.2 and Grok fast 4 (not 4.1) did the same

[half geode]

Ohhh, you mean similar outputs for same query. Yeah

[paper meadow]

People say it was the same before GLM 5 and new attention, but this really sucks. Like I had 5 attempts of it creating a name for sci-fi android NPC in reasoning trace and all 5 times it was ARIA-7 or smth close to it. With Temperature close to 1

[half geode]

Yeah I noticed it in RP at temp 1

[paper meadow]

And it always starts 1st sentence/paragrapgh the same, only differing further in answer, like it's low top_k filtering options and paths in the beginning

[paper meadow]

So it's not because new attention, and was the same in 4.5-4.7?

[half geode]

Hmm? I mean just in 5

[paper meadow]

So it was introduced in 5, not before that

[half geode]

Not sure, I think so?

[mossy charm]

its their chat client its not good

it has this weird minor problem with not using markdown right sometimes too

See youre not supposed to use bullet point in markdown

you use *

[paper meadow]

I use through api. Other from stubborness in same options over and over, I have no major complaints

[chilly leaf]

temp 1 was never enough for me for 4.x, I had to bump it to 1.4 with 0.02 min-p to be semi acceptable

[paper meadow]

Very few providers support min_p nowadays

I can't go higher than 1.1, and it doesn't change much, 1.2+ breaks everything

[chilly leaf]

dang

more providers should add min_p and at least 2 temp max, min_p is such a powerful sampler

[paper meadow]

I think there were more before, and even top_a too

[wooden briar]

Got this for the first time using z.ai. It's totally fair, I'm a free user and can switch to API... looks like they're trying to make things more stable for paid/api users

AIn't even mad

I've gotten SO much free usage out of them

[half geode]

@placid crescent It's not too late for you to be the #2 GLM fan and I'll be the #1 GLM fan

[exotic quarry]

use from kilocode , its free for some days

[worn yarrow]

they sent out an apology

[wooden briar]

Class

[ornate root]

I'm not getting any GLM5 responses on OC at all, and I'm on the coding plan; apparently they still don't support the Lite plan. It's been hammered since it came out.

[half geode]

So have I 🍻

But yeah, GLM and Kimi have both been getting slammed. Kimi for OpenClaw and GLM likely because it was free in Kilo / OpenCode.

Looking like the street lights in a US traffic jam

[halcyon estuary]

?? provider SiliconFlow

[obtuse spear]

i get this with g3p sometimes

[tall wasp]

does anyone feel like the fireworks provided glm 5 (and kimi k2.5) feel a lot worse at tool calls and overall quality recently?

edit: nvm, have actually been routing to nebius, fireworks seem to have heavy rate limiting because of openclaw

[wooden briar]

https://docs.z.ai/guides/overview/pricing

GLM 5 Code model appeared in pricing

[half geode]

More expensive =(

But will probably be dope

[viral bone]

for large roleplay, dont work...

[paper meadow]

Explain

[cobalt furnace]

For large work, doesn't roleplay....

[ornate root]

... and it's even more expensive than GLM-5! I guess things must be going well for them.

[paper meadow]

All work and no roleplay makes GLM a sad boy

[half geode]

GLM gets a lot of roleplay =P

[paper meadow]

https://tenor.com/view/starship-troopers-gif-18102989

[umbral stag]

Mostly mean long context

[umbral stag]

How many tokens is it?

[paper meadow]

It should be better compared to most models

[half canopy]

What settings do you use for role play? Do you use thinking or non-thinkign

[paper meadow]

It seems to need higher temp than average LLM

I can't get non-thinking from my source

For non-multilanguage, I can get 1.1 temp and 1 top p

[half canopy]

Is it better than deepseek v3.2? I mean is it worth paying 6x?

[paper meadow]

No way

Maybe could be better, but not 6x better

If even

[half canopy]

I see , thanks for replying

have you tried any other models like the minimax models?
Which one do you think is the best for RP performance and price wise.
IK opus/sonnet are kings but too expensive. Also I don't care about censorship since i never do anything sus.

[ebon depot]

that new aion model isn't bad as long as you're not doing anything too complex

[paper meadow]

Minimax is bad, Kimi 07 and Kimi 09 for <16k roleplay, Kimi K2.5 thinking after 16k

[half geode]

Wait, GLM is too expensive but you're using reasoning in Kimi?

[paper meadow]

Kimi's output is not that expensive, plus potential cache hits from moonshot and novita (?), plus you can't have cheap long context convo without reasoning models. Having first XX turns being written by original Kimi models helps to pick up style by K 2.5

I would also advice model hopping but this is like rocket jumping == advanced practice

[half geode]

Not that its reasoning is specifically expensive, but I'd be surprised if at 16K-32K Kimi reasoning was cheaper than GLM non

[paper meadow]

At this context input is eating a lot of price share compared to output. But also Kimi K2.5 is reliably better due to being bigger with shared and activated parameters. And does not suffer from structural repetition

[half geode]

Interesting. I have done the most comparisons in a kind of roleplay assistant mode, but maybe I should do more in actual RP. GLM might be biasing me in that sense because it feels the most human by a LOT

[paper meadow]

I have my own personal tests evaluations so my opinions could be not only based, but biased as well

[umbral stag]

So i doing some detective role-play, where it taking reference from real life with alternative path

At some point i indulge in case about gang where it referencing china or hongkong gang, it's the stop and tell me it couldn't provide it, seems like the blocked is because it's prohibited content lol

But with russian, italian and japanese gang/mafia/yakuza it doing fine

[half geode]

Web UI?

[umbral stag]

Using GLM Chat frontend

Direct chat, No API

It's Fair imo

[half geode]

Yeah, the web UI for all Chinese models have post-request censorship

[umbral stag]

As long as the model which being serving by other than Zai aren't have that censorship

[half geode]

Hmm? Kimi does

Oh, no, their API is fine

Just not webui

[viral bone]

nop, its better deepseek.

[plain topaz]

Glm 5 code this week🙏

[formal epoch]

美女

[arctic portal]

its surely far better than deepseek but not 6x

its the best open source roleplay model

i would say its 2.5x better than deepseek

k2.5 is just derpy and weird for me

glm 5 has a slight positivity bias

but its realistic

[halcyon estuary]

for some reason deepseek 3.2 is doing so good for some tasks

it has good portuguese knowledge and it's dirty cheap

only thing is it can't make tool calls reliably

[ancient flare]

you mean opus 4.5 biased?

glm 5 is amazing because it btw, even if it got way more censored

[arctic portal]

yeaaa opus is so positivity biased

i have to bring kimi sometimes to make it do grim stuff

it just cant do dominant/dark scenarios

kimi is kinda like r1 0528

very unhinged, negatively biased and kinky

[ancient flare]

if it wasn't dumb (even their thinking model), it would have been sovl

[cyan sequoia]

Yea. GLM 5 is like opus that knows less while kimi is like gemini that is even crazier.

[half geode]

没毛病

[uncut quest]

I heard that GLM-5 is more censored than it was during Pony Alpha. Just how censored is it for roleplay (if at all), if anyone knows?

[cyan sequoia]

Not at all

Never had it reject anything before

so no idea why some people say its censored

[uncut quest]

Awesome, thanks for the clarification. Was probably gonna be putting some money into it soon, so I was wondering if the censorship was true or not

[cyan sequoia]

Imo for creative writing its only 2nd to opus itself

[paper meadow]

Not with that rigid stubborn structure

I can't get 2 different responses from it

[umbral stag]

It seems they done a bit more tuning after the pony, but i could be wrong.
Seems a bit strong if you have some injection to break it so it become like the most evil being in the world

Clear comparison that i have is with their older model label GLM4.5, making GLM4.5 the most evil being is much more easier

[ruby widget]

Too long and too censored. But I will disclose im using it through openrouter/chub ai, which could have some filters

[cyan sequoia]

Ive literally never had it reject anything

even dark very nsfw stuff

[half geode]

Yeah, sounds like prompting issue. Model is not censored, and is fantastic at RP

[arctic portal]

it will maybe refuse every 1 in 100 prompts during extremely dark rp

[quiet violet]

https://x.com/basetenco/status/2029740408419586522

pls add baseten :3

Baseten (@basetenco)

We've launched the fastest GLM 5 API available at 190 TPS and 0.79 sec TTFT with the Baseten Inference Stack.

Ready for your coding and agentic workflows.

https://t.co/iiRmQK3D5U

▶ Play video

X · Mar 6, 2026, 2:06 AM

[mossy charm]

FYI if anyone is sick of the glm chat client eating your prompts - I made one for chat.

[worn yarrow]

Do tell

[mossy charm]

Oh I didn't want to spam the links posted it in #app-showcase

zoltun.org or zoltun-org github tho

[half geode]

Idk how they don't even have an app. Coding focus I guess

[mossy charm]

Who z.ai? Their web client is kinda ... eh

[half geode]

Yeah. I can't even toggle off thinking on the mobile site

But no native app whatsoever. I guess Qwen and Daobao don't either, although maybe just a China thing so I can't see it (?)

[mossy charm]

Yeah mine, thinking is off by default

you can change it in settings.json along with the endpoints tho

[mossy charm]

Alright the CLI zai check has a 1.0 release now too (on the right) if you want to check your usage from CLI
https://github.com/cioran0/zai_checkbalance

OR does a better job tracking than their own API does lol

[mossy charm]

They actually have an app it just sucks. IDK if its in English either

[half geode]

Also just to keep glazing this model, I think it has to be the most human-like. Very enjoyable to talk to.

[light dust]

It says no sometimes at times that other models would say yes

In a good way

[half geode]

Interesting, because in terms of benchmarks the biggest problem I've seen with it is that it isn't very assertive and is unlikely to push back on nonsense

[paper meadow]

I tried all options to make it output variable texts, but no

[half geode]

What does variable outputs have to do with it sounding human?

[paper meadow]

When it starts each message the same way and uses same structure, i lose any suspension of disbelief

[half geode]

I find the consistency between attempts at the same output to be even more human. Structure between subsequent messages, sure, but every model does that. I haven't found 5 to do it to a noticeable level when it's in just chat mode

[light dust]

Eh not per se on nonesense but more some controversial takes on obvious yes’s

Which make you think in a new perspective

I appreciate that

[half geode]

Ah, gotcha, interesting!

[obtuse spear]

i wish it was a bit cheaper

[high dove]

https://openrouter.ai/z-ai/glm-5-turbo/api

[primal thorn]

what is this?

The price is the same, so it does not seem like a mini model

I am confused

[umbral stag]

Faster endpoint?

It's actually more expensive than the normal version, right now is still the discount phase it seems

Normal 3.2$ 1M Output | Turbo 4$ 1M Output

[inland heath]

Weird to call it a "new model" if it's just a fast endpoint 🤔

[static pond]

nothing on z.ai about it that i can see

[umbral stag]

Could bit a bit more tune for openclaw
Could it be targeting openclaw market specifically?

[static pond]

Could it be GLM-5-Code? https://docs.z.ai/guides/overview/pricing

[high dove]

https://x.com/ZixuanLi_/status/2033213795296752106?s=20

Zixuan Li (@ZixuanLi_)

New experiment

X · Mar 15, 2026, 4:08 PM

[ancient flare]

should be under "z.ai turbo" instead of new endpoint

[halcyon estuary]

but it looks like its been trained for these agentic stuff again?

[obtuse spear]

yeah it seems like an actually different model instead of just existing but faster

at least thats what the description says

[high dove]

https://x.com/Zai_org/status/2033221440267280696?s=20

Z.ai (@Zai_org)

Note: As an experimental version, GLM-5-Turbo is currently closed-source. All capabilities and findings will be incorporated into our next open-source model release.

X · Mar 15, 2026, 4:39 PM

https://x.com/louszbd/status/2033224682565161461?s=20

Lou (@louszbd)

pony-alpha-2 has finally leveled up into GLM-5-Turbo
can’t wait to see how it performs!

DM me with your User ID if you need a rate limit increase for GLM-5-Turbo
https://t.co/aNuac8wLA4

X · Mar 15, 2026, 4:51 PM

Seems like its likely a unique model

[heavy rose]

is this a different model ?

i don’t see any model weights on hf

[exotic tinsel]

@heavy rose

[heavy rose]

ah ok

[livid grove]

seems glm 5 turbo faster than glm5🫡

[cyan sequoia]

Probably GLM with multi token prediction layer like how qwen did it and other speed ups

[mossy charm]

extra openclaw related

[untold remnant]

GLM 5 struggling to count

[untold remnant]

bruh

is GLM 5 ok?

[half geode]

Anti-GLM psyop

[untold remnant]

It was working great for me and then the last couple days I just see it struggling

[mental belfry]

the tok/s is so bad and its been ages

[obtuse spear]

that’s what happens when you pick a wicked attention

[fresh osprey]

So many rate limits

[reef stirrup]

Has anyone noticed a loss of quality the past week or so?

[solid silo]

5.1 wen

I've used the GLM coding plan, and the Opencode Zen, honestly the GLM coding plan gives way worse responses

i believe Opencode Zen uses fireworks provider, so the Openrouter GLM 5 with Fireworks provider should be good

or Baseten, since i see it has better uptime

[formal sapphire]

Perhaps too many people are using coding plan，z.ai can't afford that many requests

[solid silo]

i believe they also route the requests to a subpar version with quantization

[stuck stag]

Yep :(

[untold remnant]

Their discord is full of people complaining about the coding plan, their glm 5 is completely busted, supposedly other providers don't have these issues. And they just tell you to use turbo instead of providing any explanation or even really acknowledging that something is wrong and they're fixing it lol

[half geode]

Yeah, coding plan is borked

Use OR or Opencode's plan

Sad, really cool org but they kind of fucked themselves on PR with the coding plan stuff. For some reason they just rolled it out to Lite plans despite all the usage issues?

[untold remnant]

I heard it got revoked from lite plans lmao

[viral bone]

anynone have this error????

Quota exhausted, please check your API provider account v0.77
openrouter: z-ai/glm-5

{
"error": {
"message": "Provider returned error",
"code": 429,
"metadata": {
"raw": "z-ai/glm-5 is temporarily rate-limited upstream. Please retry shortly, or add your own key to accumulate your rate limits: https://openrouter.ai/settings/integrations",
"provider_name": "DeepInfra",
"is_byok": false
}
},
"user_id": "

[light dust]

Suffering from success

[manic timber]

Will there be another free glm model or just the 4.5 Air?

[ancient flare]

glm 4.7 flash is better and free on their official api

[manic timber]

But how do you use it in Chub?

[ancient flare]

idk, i use sillytavern (while importing chub cards)