tawny raft
MiniMax-M1
#MiniMax-M1
wooden quail
Benchmarks make Minimax look like its on-par with Gemini 2.5 pro
I wonder if it token spams, though
analog steeple
their previous model minimax-01 was also good and very underrated imo
came out at the wrong time, like a few days after deepseek-r1 and wasn't a reasoner, but it had really strong long context.
misty onyx
any plans to host it?
wooden quail
any plans to host it?
Open Router doesn't host models. You have to wait for providers to host it
obsidian yoke
anyone got some feedback on how it RPs?
fallow goblet
@twilit jacinth @shadow tide
Do we still have deals with MiniMax?
shadow tide
it’ll come online today
round mantle
anyone got some feedback on how it RPs?
bump
fallow goblet
anyone got some feedback on how it RPs?
Not yet being hosted
dusty kettle
What’s the TPS look like?
analog steeple
Minimax themselves are almost certainly hosting it
shadow tide
they are, we'll have it today
humble talon
waxen lily
How's that coming along
waxen lily
they are, we'll have it today
The hosting
Cuz the demo asks for signing up or smth
And it's all chinese and chrome's auto translate isn't working
obsidian yoke
waxen lily
ooohh
cool cool thanks
the reasoniing looks solid
so far only reasoning...
hope iit doesn't do a qwen
lord wiill it ever stop thinking
the pygame app crashed at first cuz it used the same identifier to refer to different things
after fixing it it looks kind of boring having asked it to be creative
somber sapphire
🫠
elfin robin
fallow goblet
Okey, this model actually quite good
Has been tested it, not as smart as other big model but it didnt have the problematic way of the other model.
wary vine
unfortunately only the novita ai's endpoint supports tools :/
fallow goblet
How do we get pass limit of 1000 thinking token? i alwasy get that but i want to try and get 80K token thinking
Is there a command on their page about it or it depends on some parameters?
hasty carbon
lord wiill it ever stop thinking
it's qwen on steroids
i reached 895s of thinking with no output
humble talon
Finally a model that just thinks
strange marlin
very verbose + ultra slow inference isn't such a great combo tbh.
Will see how this one does, though since MiniMax-01 also had grand marketing and was pretty meh overall.
worthy socket
wtf 1m context window
hasty carbon
Finally a model that just thinks
after 15 minutes of pondering, the model decides the best course of action is to remain silent.
humble talon
the smartest llm would know better than to interact with humans
worthy socket
Finally a model that just thinks
LMAO
viscid nexus
Is this the o3 pro experience? 17 mins is 0_0
supple estuary
What is "extended" it's more expensive, but says smaller input context. Longer output??
strange marlin
asking a simple 3-step math question 😅
wary vine
wait!...
obsidian yoke
model be like
<think>
So the user wants me to roleplay them as Sasuke in Naruto. To achieve peak immersion, I must emulate post-trauma stoicism, pre-revenge arrogance, mid-revenge angst, and subtle homoerotic tension with Naruto, while maintaining canonical speech brevity and ignoring filler contradictions. Compute precise balance of “Hn” frequency. Cross-check clan massacre references every third sentence. Initiate brooding glare subroutine. Ready.</think>
Sasuke: Konnichiwa
stiff hill
model be like
<think>
So the user wants me to roleplay them as Sasuke in Naruto....
Banger
stiff hill
wtf 1m context window
Yeah but they had 4m context window model previously and it was doo doo
Large context window really a marketing gimmick nowadays
fresh prism
This model is peak
fallow goblet
asking a simple 3-step math question 😅
Haha
They truely using manually calculation
cunning flint
asking a simple 3-step math question 😅
The best part is that it isn't even right
hasty carbon
I wanted to test out MiniMax's older, non-reasoning model, but it seems like it just instantly Internal Server Error's
strange marlin
Tested MiniMax-M1:
At 456B too large to run local, and as a ultra-verbose reasoning model and slow inference speed via API, found this model to be unusable for any real work.
With 92/8 reasoning split, this model spent most of its time thinking, sometimes exhausting all 40k max tokens without giving a single reply token.
In terms of capability, I found it to be competent at my tech and coding tasks, while producing fairly average results in other areas; around Qwen2.5 Max level.
I place this model in the same category as Phi-4-reasoning-plus or, to an extend, Mistral Magistral, not really usable. But, YMMV!
hasty carbon
**Tested MiniMax-M1**:
At 456B too large to run local, and as a ultra-verbose re...
I sorta agree with this, however I believe it's around qwen 3 235b, not phi-4 or magistral.
It gets into more thinking loops than qwen, but when it doesn't it's around on-par with it when it comes to code quality
fallow goblet
Finally we have model that can goes above 192K with 71% understandbility
Minimax really are different, hope they can keep on going improving their model
Aint the smartest yet but really competent model to fight the gemini 1M context
civic coral
Can you link this table? Does it go up to 1M?
wooden quail
Can you link this table? Does it go up to 1M?
It doesn't go up to 1m
This graph is from fiction livebench: https://fiction.live/stories/Fiction-liveBench-Mar-25-2025/oQdzQvKHw8JyXbN87
strange marlin
Finally we have model that can goes above 192K with 71% understandbility
the model is ultra slow though. 22tok/s currently at long cot reasoning, when adjusting for verbosity, its effectively the same speed as claude generating at 1.6 tokens per second, unusable.
just for reference, when I set up some chess matches (usually takes a few minutes), I had to have it run all night > was still playing > and when I then came home from work almost 24 hours after starting some matches finally concluded. that is the degree of usability we are talking about here.
serene laurel
It doesn't go up to 1m
This graph is from fiction livebench: https://fiction.li...
This is 40K model not 80K
fallow goblet
the model is ultra slow though. 22tok/s currently at long cot reasoning, when ad...
Still impressive
No other model than google model has able to achive good context understanding at high token amount, maveric still below them and they are much more uncensord unless if we talking about china stuff.
wooden quail
This is 40K model not 80K
Context length, I meant
wary vine
maverick is just unusable in every way for me, very awful
weird style of writing, forget most stuff so awful for summarizing things like youtube videos, doesn't know what to prioritize
serene laurel
A new provider for hosting minimax m1
https://www.siliconflow.com/models/minimaxai-minimax-m1-80k
fallow goblet
A new provider for hosting minimax m1
https://www.siliconflow.com/models/minimax...
@shadow tide @twilit jacinth
You guys need to contact and work with em
