Xiaomi MiMo V2 Flash

[viscid fog]

https://openrouter.ai/xiaomi/mimo-v2-flash:free

[hardy plume]

is this running on ascend chips?

[hearty shale]

🔥

[short summit]

@viscid fog using that models :free counting to 1000 requests per day limit on 10$+ user plan?

what if there is only :free model variant and no paid? We can't use it after reaching 1000 RPD?

<@&1384697330254610442>

[hardy plume]

yes

[viscid fog]

the Xiaomi free endpoint has unlocked RPD limits

[short summit]

@viscid fog Nemotron 3 Nano 30B A3B also have unlocked RPD limits? this model also don't have for now paid variants

[viscid fog]

that model also does not have RPD

[short summit]

okay thanks ❤️

[pastel turret]

Is this endpoint being trained on? Or just free, but no training?

[quiet onyx]

weights: https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash

[honest oxide]

Prompt: Write a short story about a deal with the djinn gone awry
Output:

[viscid fog]

free & not training

[pastel turret]

Huh this model seems pretty good

[honest oxide]

I actually really like the writing

[pastel turret]

I like the coding style

Much more than the default style of like GPT 5

And it seems to be coding pretty well too

[honest oxide]

gonna use this to co-write for a bit and I'll report back if I find the writing to get annoying

[quiet onyx]

blog mentions this being under the MIT license which is great but there is no licence in the repo or if its "modified MIT"

[pastel turret]

if hybrid attention like this can be widely adopted + competitive it'll be so sick

110 TPS on a top model is massive

[honest oxide]

anyway this model is good at writing and free so

enjoy it before jai finds it

[round cradle]

Looks like there's a moderation layer

It will reject 'high risk' prompts

[honest oxide]

ah

[round cradle]

Which in my testing is a pretty wide category :/

[pastel turret]

all rules and UI work perfectly

castling, en passant, promotion, etc.

[quiet onyx]

this model scored 40% on my spatial reasoning test from a 20 year old children's medieval fantasy game

[honest oxide]

😭

[pastel turret]

did very well in a personal bench, just as well as gemini 2.5 flash / grok 4 fast / deepseek 3.2

[honest oxide]

what's the API pricing? It might replace grok 4 fast for me

[pastel turret]

no idea, but I'd guess cheaper

[honest oxide]

also 15 day free period ⁉️

[pastel turret]

probably figuring out pricing

hybrid models are probably harder to price correctly than traditional models

idk

[honest oxide]

holy

[pastel turret]

based

not AMAZING at terminal bench, but not awful?

GLM 4.6 scores 24.5%

this is actually the second highest open model on terminal bench perhaps?

yeah I think so

deepseek is higher but has other factors that make quite bad to use for coding, like insane hallucination

This model is still prone to hallucinations, but vibes seem better on its niche knowledge than most other open models I've tried

[quiet onyx]

blog mentions this model is under MIT licence but Github has Apache???

https://github.com/XiaomiMiMo/MiMo-V2-Flash/blob/main/LICENSE

[brittle notch]

Improvement

[pastel turret]

this model might actually be the best flash level model out there rn

[honest oxide]

this is 100% taking the role of grok 4 fast for me

[pastel turret]

scoring higher than grok code fast on terminal bench

[hardy plume]

sounds like a great model, now i gotta try it

[pastel turret]

and if it wasn't apparent: scoring higher than minimax, glm 4.6, kimi k2 (and thinking variant), and even Claude 4.5 Haiku

[hardy plume]

@viscid fog

maybe because of the word "brutalist"?

nope.. not because of brutalist

🤔

maybe my system prompt

[honest oxide]

oh right refusals

hopefully it doesn't do that when I try to use it to answer tool calls?

[pastel turret]

sadly, this model does not have up to date knowledge it seems (e.g React Router V7, Tailwind V4)

[hardy plume]

i have no clue what in my system prompt could be causing this

k time to cut out parts till it works

[honest oxide]

looks like .8 temp and .95 top_p is recommended?

[quiet onyx]

where is this from? I couldn't find any pricing info on their official API docs https://platform.xiaomimimo.com/#/docs/pricing

[honest oxide]

https://platform.xiaomimimo.com/#/docs/quick-start/first-api-call

oh

you meant the pricing

https://platform.xiaomimimo.com/#/docs/news/news20251216

[hardy plume]

okay it seems to be this causing it

whatever i can omit that to use it anyway

[quiet onyx]

$0.1 per million input is low but not as low as like $0.02 from some other models if it doesn't support caching

[honest oxide]

tbf it doesn't say

[hardy plume]

and somehow only sometimes

[honest oxide]

oh my god

this model is actually really good at being agentic

thank you xiaomi

[visual creek]

SO GOOD

[quiet onyx]

atleast they didn't commit any chart crimes here

[brittle notch]

did retry, stuck in reasoning

[steel harbor]

its a good model, sir

[brittle notch]

ohhh, the

You are MiMo-V2-Flash (free), a large language model from xiaomi.

Formatting Rules:
- Use Markdown for lists, tables, and styling.
- Use ```code fences``` for all code blocks.
- Format file names, paths, and function names with `inline code` backticks.
- **For all mathematical expressions, you must use dollar-sign delimiters. Use $...$ for inline math and $$...$$ for block math. Do not use (...) or [...] delimiters.**

fucks with it's head

infinite headpats. this model gets it to not be awkward

so it is VERY system prompt sensitive and gets hyper focused on it's objective to the point of getting confused

[honest oxide]

guys bad news, I just saw a screenshot from the jai server

[hardy plume]

safety filter should help if its this sensitive

[honest oxide]

safety filters save us

[brittle notch]

sir, you want to nurf an already silly model?

https://tenor.com/view/my-hero-acadamia-deku-excited-head-desk-wiggling-gif-5497470

[hardy plume]

well its open weights, providers will probs pick it up once they resolve the license thing

[honest oxide]

I like this model for summarizing

[quiet onyx]

I made a github issue about that https://github.com/XiaomiMiMo/MiMo-V2-Flash/issues/2

[honest oxide]

overall very happy with it

[viscid fog]

I am fixing this now

[hardy plume]

based on their gh these are very recommended

[visual creek]

Hightly recommended, even

[hardy plume]

also seems to be a interleaved thinker

like week number maybe?

[visual creek]

Maybe it should be {month} ?

[hardy plume]

i asked gemini 3 to translate the system prompt

thats probably what it would come out to

[quiet onyx]

so far this model have been performing my SWE tests competently, such as compiling program from the year 2000 on modern toolchain https://gist.github.com/kth8/0897f24ce7c7bed643291dd6ff658e15

[brittle notch]

my friend sent this, very weird result for multiplication

[hardy plume]

talk about optimizing, inlining C

if only it worked that easily

[pastel turret]

I'll be extremely pleasantly surprised if this ends up to be the best open agentic coding model

and it seems like it may be that

(deepseek 3.2 excluded for being too spiky, slow, and full of hallucinations)

[visual creek]

I guess Cursor Composer V2 will just happen to drop a week from now as well then eh

[pastel turret]

I was about to say

this is like a better composer

lmao

[visual creek]

I think composer is just GLM 4.5-4.6 finetuned no?

[pastel turret]

if they do their postrain regime on this model instead of GLM 4.6 (assuming that's what they're using)

yea

I think so

lines up with the pricing

they could run this on GPUs instead of cerebras or groq or whoever, and serve it WAY cheaper

one composer's issues has been it's surprisingly expensive, I assume because it's on cerebras/groq to go fast

[steel harbor]

is this actually securely better than glm? im having mixed vibes with it

[pastel turret]

it seems better to me

maybe not design sense

[left tree]

If it is flash then the pro model will also be on the way.

[brittle notch]

doesn't seem to have real understanding on stuff and seems to be trained on best Q&A from our short test, but it is adorable

did not test the agentic coding of it, so can't prove or deny bulbasaur

[quiet onyx]

Honestly shocked Xiaomi coming in out of nowhere and making a model this good at agentic tasks

[honest oxide]

I am very pleasantly impressed by it

[pastel turret]

Did better or on par in a couple vibes tests, but the big thing to me is it’s performing significantly better on TerminalBench 2.0

Which imo is a quite high quality bench

[brittle notch]

no, aider benchmark is imo, but no unofficial results yet

[umbral oxide]

Ok Xiaomi, I wasn't familiar with your game...

Looks interesting

[pastel turret]

I think aider is not relevant anymore, it doesn’t test tool calling or anything

• it’s been around too long, data is definitely in training sets

The benchmark hasn’t even been updated with latest models in over a month it seems like

[brittle notch]

in my tests, it is still representative of the model state. i don't think it is a benchmark they can saturate.

[brittle notch]

ppl in discord testing themselves

[pastel turret]

[jovial plinth]

😭 this is so good guys.

[pastel turret]

seems pretty good, testing it with Opencode on a large codebase

not "amazing" or anything

like composer 1

[rustic river]

I'm constantly getting

421 {"error":{"code":"421","message":"Moderation Block","param":"The request was rejected because it was considered high risk","type":"content_filter"}}

even if I just say "hello"

Not sure how you guys have managed to get it working in your agentic coding tools

[hardy plume]

hey indeed.

jeez this model is quite good, no longer getting the safety warning atleast, managed to add a codex-like session system to my cli first try, also went around my harness by using shell commands to read files because it didnt like that i didnt have any line range support for reading.

this model is quite eager to write test python scripts, then use them to test, very practical and seems to actually delete them after too

similar to claude

[pastel turret]

First Token Latency is the only issue I have with this model currently

it's ~2.5s on average, which cuts down significantly on the benefits of TPS

if it was like 500ms like other models it would be amazing

[hardy plume]

probably due to the region and their filter

[quiet onyx]

One limitation I found with this model is it can only make 1 tool call per turn. That became really inefficient and troublesome here when I asked to setup a whole cluster and it can only run commands or write file to 1 machine at a time https://gist.github.com/kth8/f2d17b3b8b017055a4daedd03994d2f6

I've seen Grok in comparison make 5-10 tool calls at once per turn to manage multiple machine in parallel

[pastel turret]

It seems like it’s doing parallel tool calls in OpenCode… are you sure it can’t?

[quiet onyx]

I haven't seen it. Is there some magic phrase I need to put in the system prompt for it to do it?

[last sluice]

Is this model permanently free, or just for a week or so to test, like Grok 4.1 was?

[quiet onyx]

free time left: https://platform.xiaomimimo.com/#/docs/pricing

[last sluice]

Perfect, thank you.

[thick niche]

o z o n e

[lean maple]

Xaomi has been throwing 500s and 524s

Its not common but it happens

[hardy plume]

this model is a good replacement for grok code fast if providers will step that low on price & have caching

grok code fast is good but its tool calls are hit or miss, sometimes it tries to call them in reasoning and says it did do the changes but didn't

[lean maple]

xiaomi mimo is unsuable now due to rate limits

[pastel turret]

@viscid fog btw Novita is hosting this model now, can we get a paid endpoint?

[bleak forum]

https://x.com/XiaomiMiMo/status/2002401028722106843

XiaomiMiMo (@XiaomiMiMo)

MiMo-V2-Flash scores 66 on the @ArtificialAnlys Intelligence Index — #2 among open-source models and #8 overall! 🎉🎉🎉

Designed for Agentic AI — now with the benchmarks to prove it: #1 on τ²-Bench Telecom for agentic tool-use among all evaluated models. ⚡⚡⚡

Frontier

X · Dec 20, 2025, 3:29 PM

[proven sequoia]

did anyone manage to get this working with interleaved thinking in agentic coding tool?

[naive hedge]

thinking is disabled in their anthropic API for some reason, and can't be turned on

[buoyant phoenix]

not doing full testing, but as proxy, not great at chess, bottom 15%

[proven sequoia]

ah... that might explain when I tried to do integrated thinking with OpenCode via OR API it just put the function call inside the think and printed the XML instead of interceptin it as a function call

but their docs say it supports tool integrated thinking

wierd

[naive hedge]

I am getting that too in opencode, but that's slightly unrelated to Anthropic API as OR almost certainly is using OpenAI API format where thinking can be enabled. But not sure what's up with these XMLs here.

OpenAI spec doesn't really support interleaved thinking (I don't even think there is a spec?), iirc opencode conditionally turns it on for a few models such as new deepseek 3.2, probably not for this model yet anyway

[proven sequoia]

DeepSeek API does it in its own adapted way, so does OpenRouter

however I was hoping that it would work using a model via OpenRouter, since the API to the app is the same

[pastel turret]

I quite appreciate the coding style of this model

unlike something like gpt 5 which I still hate the coding style of

this model structures code well and doesn't have weird stuff mixed in like gpt would (e.g if clauses with like 5 && conditionals to validate the type of a parameter)

[vast mirage]

Not a really useful topic. But this model is good for Rp too

[strange bison]

Compared to what models?
Claude or other models like deepseek?

[vast mirage]

It is a bit smarter than deepseek. I havent use claude since i aint paying allat for claude. So this is being compared to v3.2 exp, v3.2, grok 4.1 fast

so most of the nsfw models that is normally used in RP.

Nah i take my words back

it is yet not on the level of v3.2

[honest oxide]

I like deepseek better

[vast mirage]

yep, deepseek is better

[stone tangle]

We've seen a ton of messages above about moderation. You can't really roleplay with constant moderation errors.

[vast mirage]

huh?. i never got moderation message at all while i am mimo. even for nsfw rp

[stone tangle]

Maybe not a "ton," but if you scroll up, you'll see a couple.

[strange bison]

Bruh. What kind of role plays are you guys doing that you get constant moderation errors. Genuinely curious

[naive hedge]

In most recent opencode, it seems you can add this to a model:

"interleaved": {
  "field": "reasoning_content"
},

seems to work here with OR (no more xml errors), though this model tends to get loopy kinda soon

[proven sequoia]

oh thankyou, I will try it again later

[silver mortar]

this is sooo fast

[rustic river]

The model has been glitching in opencode today with broken outputs and premature stops

[proven sequoia]

This model seems pretty smart but I can't get it working properly in opencode

[ionic surge]

whats weird to me is the strong recommendation to turn off thimking for agentic stuff

i dont really understand WHY

also the promise of them not logging prompts i kinda doubt

[proven sequoia]

just how it was trained I guess, although I don't know why it also supports interleaved thinking with that being the case

[ionic surge]

kinda interesting how no providers have launched support for this model

on a paid endpoint

and idk when xiaomi will end this (if they will?)

@viscid fog do you know anything about this?

[quiet onyx]

free time left: https://platform.xiaomimimo.com/#/docs/pricing

[strange bison]

I know providers like chutes do provide this model , but open router hasn't offered the paid version yet

[proven sequoia]

Xiaomi will presumably swap their endpoint to a paid one soon. The model itself seems really smart (I went through a difficult problem with it in the chat interface), but interleaved not working properly in opencode for me right now.

this was using the "interleaved": {"field": "reasoning_details"} trick but maybe there is some other stuff that needs to be done for it to work properly.

[pastel turret]

novita does, openrouter just not routing to it yet

[rustic river]

I mostly do brownfield coding, and it solved problems that GPT 5.2 Codex / Gemini 3 could not

It's remarkably persistent. One of my favorite models of 2025 so far, given that it's also fast.

[slow jay]

no paid verison yet

[strange bison]

What parameters are you using?

[hardy plume]

they extended the free access

[pastel turret]

We got a paid endpoint but OR hasn’t put it up yet..

(Novita)

[rustic river]

Default!

[slow jay]

how much more

[hardy plume]

till jan 20

[slow jay]

and then what pricing?

[quiet onyx]

https://platform.xiaomimimo.com/#/docs/news/beta-free

[fickle mortar]

@viscid fog Novita AI has support for the paid mimo v2 flash, can we please get support on the openrouter gateway

[fickle mortar]

i need to use it in a commercial application

[strange bison]

The Mimo model has gotten a new snapshot 🥂

The free tier continues

[ionic surge]

yay

they finally say that the thinking mode is recommended

[slow jay]

till when?

[clear ember]

why is this model getting deprecated?

[viscid fog]

the free model is

[strange bison]

26th Jan, I assume

[clear ember]

so it will be paid only model? and why?

[lean maple]

Yes, it will only be paid

The provider that's providing the model for free will stop providing it for free soon

[quiet onyx]

abuse it while you still can

[ionic surge]

dont have any workloads to abuse it with 🥀

[hoary quail]

does anyone here know which AP prompt works best with mimo v2? for RP specifically?

[plain coral]

I have a question about the deprecation of this model. I tested it over the last 30 days very extensively and find it very useful for my agentic coding tasks. Other providers hosting this model too and I tried it out on them and have very different behaviors. How do I know that they run the same latest snapshot? Should I ask them all? And what is the latest snapshot at all hosted on OR for the free model?
Except for the thinking loop and leaking tool calls into assistant messages and thinking tokens, the model performs very well.
At least, the paid providers let me choose the seed parameter.

[vast mirage]

Anyone got recommendations for free models that is on same level as mimo v2

[hoary quail]

unfortunately not really. mimo v2 with titi's prompt is probably as close u can get to a decent 3.2 experience as you can right now
and its the cheapest option right now too

[vast mirage]

yea I think i might switch to v3.2 with a provider that allows for cache read

[analog juniper]

This model is a neat lil guy. I like him, especially for the price.

[pure garnet]

Oh, huh 🤔

$0.09/M input tokens
$0.29/M output token
That price is actually pretty low, around which models do you think this performs?

[hoary quail]

although personally, mimo has a 2nd person POV issue
for ideal results you need to generate 1-3 messages with DS 3.2 first, and then you can leapfrog to mimo v2 for a better stable experience that's different from 3.2 and cheaper

i've done extensive testing (like 200 msgs) worth, so give it a try

[hoary quail]

alo another major weakness of mimo v2 is that it really struggles to progress the scene/RP on its own
you have to explicitly prompt it (and that's even with a comprehensive AP prompt) as well

[ionic surge]

grok 4.1 fast

and cheaper than it too

and i already thought that 4.1 fast was the performance/price goat

mimo takes it

[hoary quail]

is grok 4.1 similar to deepseek 3.2 or something in the RP it generates?

[honest oxide]

no, under no circumstances should you use grok fast for rp

it barely speaks english

[hoary quail]

aw thats a damn shame. than any other recommendations for something thats as cheap as mimo v2?
i just need something thats on mimo v2 level but actually can advance the plot story on its own
cuz basically all i use right now is DS 3.2 with heavily nerfed context to make it budget affordable

[ionic surge]

rp 🥀

idk i dont do that

mostly for classification tasks/extraction takss etc

[honest oxide]

nothing beats deepseek on price per token. GLM has a cheap sub they offer, but that's about it tbh

I'm not sure what your budget is, but if you're using something like ST I'd say maybe look into a memory extension to save on context

[hoary quail]

i'd use glm 4.7 everytime, but its damn expensive once i get past 50 msgs
i wish it had a non thinking mode - or at least didn't use as much thinking, cuz frankly, its kinda ridiculous compared to models like r1 0528
i'm not exactly broke - i just prefer as much bang for buck llm model as possible
i use j.ai , just not sure how using chat memory affects how much tokens is consumed and whatnot

[honest oxide]

look into the z.ai coding plan then, but yeah as far as PAYG goes nothing beats deepseek

(also, 4.7 does have a non thinking mode, j.ai just doesn't support it )

[quiet onyx]

GLM are hybrid reasoning models mean you can disable thinking. You can do it via code or create a custom https://openrouter.ai/docs/guides/features/presets

[vast mirage]

it hallucinates alot for me. Like many times it would just either say something irrelevant or something that is not within scenario. Interestingly My friend pointed out that this happened in chub.ai more than it happened on Janitor

i will switch to r1 0528

[hoary quail]

yea just did some further testing
its a big nothing burger model, just generates responses that don't really move forward, more circling on the spot - no matter the AP prompt
i don't know if this will change but i would suggest to avoid it for now

[pure garnet]

GLM 4.7 has a non-thinking mode

[hoary quail]

i'm aware of that. but i dunno how to easily set it up. i don't got a lick of coding knowledge

[pure garnet]

The easiest no code way would be to make a preset that forces no reasoning: https://openrouter.ai/settings/presets

[thorn bane]

What does "deprecrating Jan 26, 2026" mean ?

[open heart]

it's being unavailable (correct me if I'm wrong)

[open heart]

it is fast but not as good as gemini 2.5 for coding

[pure garnet]

2.5 Flash is over 3x the input price and over 8x the output price, though

[river lichen]

does anyone know how to integrate mimo v2 to j.ai from the xiaomi site itself? i keep getting network errors

[half iris]

does anyone know why this is always rate limited?
I can't get any agentic coding to work with this model
it consistently stops early, before making any file changes

[rain veldt]

Do you use paid or free version?

[half iris]

paid

ok, maybe this is a false alarm
I was getting rate limits yesterday
but I just spotted a problem with my opencode config (edit: deny)
giving this another test now

[half iris]

yup, it was my config
seems to be working fine now
thanks for responding @ Monkey !

[plain coral]

The issue with this model is, that it sometimes still leaks tool calls into the thinking and sometimes assistant tokens and this stops the multi-turn inference. To use this model reliably, I needed to sanitize those tokens after streaming each block and ignore the stop_reason for that.
Additional, I suspect that the paid providers are using an older snapshot of this model because it behaves very differently by each of them.
It is such a great model, but without enough transparency hard to use without those workarounds.

[rain veldt]

I don't think xiaomi site support jai

[half iris]

I happen to be using atlas cloud as a provider
Is there any way to identify which version of the model they provide?

[lean maple]

Is there any way to identify which version of the model they provide?

Nope

[river lichen]

damn, i'd thought it'll work just like the official deepseek site, just by adding & putting its own v1/chat/completions url into j.ai proxy settings like usual

[hoary quail]

nope but i notice certain providers give real subpar responses, sometimes. so i'd recommend blocking them if ur okay with that. atlas cloud is generally terrible

[quiet onyx]

I only use official provider when possible. In this case the official xiaomi/fp8 provider is also the cheapest and fastest

[half iris]

Oh I only have Atlas Cloud and Novita AI available for this model
Chutes and Xiaomi must be blocked in my OR privacy settings

So any opinions on Novita vs Atlas Cloud? 🫠

[quiet onyx]

Novita supports prompt caching which will be the biggest cost saver

[pure garnet]

Oh, wow, TIL this has caching, this is dirt cheap

[half iris]

Yup it is
And in my testing so far, it does a decent job
One real hallucination about a .gopls.toml file which isn’t a feature of gopls
Otherwise, it’s been nice

[fickle mortar]

Novita has the lowest tps(they are running the model in vllm not sglang) and the lowest e2e latency

[quiet onyx]

why are you stating to me what I posted in my screenshot?

[half iris]

Ya Xiaomi looks like the most performant provider and supports caching
But I prioritize the privacy side above everything else so not an option for me

Atlas cloud is noticeably much faster than Novita
But Novita supports caching 🤷‍♂️

[fickle mortar]

Xiaomi also has a zdr policy

[quiet onyx]

Xiaomi is not listed here unlike AtlasCloud and Novita https://openrouter.ai/docs/guides/features/zdr#zero-retention-endpoints

[fickle mortar]

You can check out on Xiaomi api page

[half iris]

Xiaomi Retained for 30 days ✓ Does not train
^^ open router docs ^^

https://platform.xiaomimimo.com/#/docs/welcome

here’s a snippet I’ve read so far

API Services. If you use the API services, we will collect your IP address and the text information you submit to analyze the relevant instructions based on the model you select and to generate the returned content. Xiaomi will not use the text content you provide for model training or any other purposes. When you use prepaid API services, we will collect your top-up information and transaction records**.**

I don’t see the number 30 show up in that page 🤷‍♂️

Seems to me like openrouter has a different reason for the privacy settings restricting Xiaomi
Or it’s a mistake?

[strange bison]

Anyone else having an issue with the xiaomi endpoint where the model thinks forever all the sudden?

Like didn't change the prompt or any of the sampling parameters, yet it's happening consistently in the last few days

[fickle mortar]

you should optimize the hyperparams as per your application

refer to xiaomi huggingface

[strange bison]

No, I think Xiaomi had an issue on their end.
It was working fine for 7-8 days, then for 1-2 days it started capping reasoning (65k reasoning tokens) occasionally, even though I didn't change anything. Now it's fine again.
Weird. Also xiaomi end point is the only end point I direct to for the API calls becuz of cache so it wasn't a provider switch either.

[celest grail]

this is awesome

i build a two agent loop so they get to a consensus with very strict guidelines

the result was pretty good

crazy value

[ionic surge]

yeah it’s peak

[fickle mortar]

I just wished we get more providers and xiaomi oss ed the latest mimo v2

[visual creek]

This is an impressively capable tool calling agent model, nothing seems to come even close at the price.

[ionic surge]

daily goat model reminder

when v3 flash ?

im so happy w tyhis model

xiaomi's a great provider too

great caching

[pure garnet]

I agree

[ionic surge]

if this model had vision i think it would be insane

because the price/performance ratio is insanely good

i wonder if grok can reclaim its price/performance crown

[honest oxide]

would you say it's better than grok fast?

[celest grail]

yes

[honest oxide]

no structured output support though 😭

[celest grail]

well, it follows instructions very well

[ionic surge]

?

really?

it says it does on orca.orb.town

i love the instruction following on this model since i know that many other flash-style models (even g3f) dont follow instructions well

[hoary quail]

I've gotten mimo V2 flash to work very well
However my only gripe that if it had a reasoning/thinking mode

That would really take it to the next level
Cuz for price to performance - it's quite good already

[ionic surge]

it does

[hoary quail]

Oh? Where is it then?

[ionic surge]

https://openrouter.ai/docs/guides/best-practices/reasoning-tokens

just set reasoning.enabled = true or set some reasonign effort and itll turn on

i dont think you can change the effort level its just off or on

[hoary quail]

How do I enable on OR?
Do I have to make a preset or something?

[ionic surge]

do you use the api or like the chatroom

[hoary quail]

I chat using proxy on j.ai
So I guess...Api I think? Sorry, I'm not a tech guy so not sure

[ionic surge]

oh, im not sure how to on janitor

try searching up how to enable reasoning on janitor

but mimo does support reasoning

[steel harbor]

https://x.com/Lentils80/status/2034207379445698593?s=20

Lentils (@Lentils80)

🚨 MiMo-V2-Pro by Xiaomi has suddenly appeared on Artificial Analysis before any official announcements.

Mentioned as a proprietary unimodal (text-only) reasoning model, it sports a 1 million context window with an Intelligence Index score of 49, sitting just below GLM-5.

X · Mar 18, 2026, 9:56 AM

[iron solstice]

Is there a way to tell which ones dont have an rpd limit?

I find that paying for anything less than opus costs me more for anything I do since doing it correctly first time = cheapest but having stuff that would be literally free to experiment with would be nice

[versed turret]

It's not free anymore

[iron solstice]

I meant in general, if there's a way to tell which ones are 1000 rpd vs not

[versed turret]

It's 1000 i think

[junior narwhal]

Based on my experience, the Mimo V2 Flash offers the best value for money right now at just $0.1/$0.3.

[iron solstice]

Jesus Christ

At least read what you're replying to