1.
77 opinionsPricing and Market Competitiveness
Many users praise the model's affordability and generous token limits as a compelling alternative to more expensive providers like Anthropic and OpenAI.
MiniMax
MiniMax M2.7
Release Date 2026-03-18205K IN · 128K OUTOpen Model(proprietary)API available
MiniMax M2.7 is a next-generation flagship model that builds on M2.5 with a self-evolving training paradigm — autonomously running over 100 rounds of scaffold optimization during training, achieving a 30% performance improvement. It is built for complex agentic workflows including Agent Teams, dynamic tool search, and elaborate productivity tasks. The model scores 56.22% on SWE-Pro (matching GPT-5.3-Codex) and 57.0% on Terminal Bench 2, demonstrating system-level comprehension. Based on a 230B sparse MoE architecture, it offers frontier performance at just $0.30 per million input tokens.
API cost
SourceOpenRouterLast updated
Real-world usage
Outside weekly top 50
685B8/31
SourceOpenRouter
Output Speed
Slowper 2,000 tokens
53tok/s
First Output 1.3stotal: 39.0s
SourceOpenRouterLast updated
User Comparisons
Lies (Hallucination)
Honest answers
100%
Correct 0Honest don't-know 3Lies 0
Response Speed
instability ±49%
First response 6.5sTotal 13.3s
32 tok/s
27.4ms per token
User Reviews
- •
I feel like the response output format is a bit unstable.
SourceAIB Inc.Last updated
Expert Performance Evaluation
Documents and office work
All documents and office work AI rankings78.3%142427.9%61.2%25.4%10.6%0.0%34.4%48.4%
Research and fact-checking
All research and fact-checking AI rankings142276.3%
Specialized fields
13241274144114531432142013.3%89.0%64.2%63.1%65.4%89.0%88.5%75.0%61.6%86.7%46.7%28.2%42.0%8.6%77.0%45.0%40.4%88.0%54.0%33.3%47.5%
See each evaluation’s detail page for its source and last updateData sources & removal requests
Public Opinion
TL;DR
Users appreciate the high value and coding capabilities of MiniMax M2.7, though significant controversy persists regarding its restrictive non-commercial license and the reliability of community-provided local quants.
523 opinions
- 38%62%PositiveNegative
- YouTube48%52%PositiveNegative
- Reddit35%65%PositiveNegative
Reaction Summary
2.
53 opinionsCoding and Agentic Performance
Users are evaluating the model's effectiveness in agentic workflows and coding tasks, comparing its reliability to industry leaders like Claude and Qwen.
3.
28 opinionsLicensing and Open Source Definitions
There is significant debate and criticism regarding the 'modified-MIT' license, with users arguing that the non-commercial restrictions disqualify it from being truly open source.
4.
32 opinionsLocal Hardware and Quantization
Discussions focus on the hardware requirements for running large quants, alongside technical frustrations regarding the quality of specific community-provided GGUF files.
Actual opinionsselected excerpts
As if we could stop waiting :) Anything open-source that costs so much money and effort is always worth the wait, and the community is as grateful as ever
From my testing glm, especially glm 5.1, is better in general. But minimax is much smaller and punches well above its weight
M2.7 is the most excited I've been for a release in a while. I've been using the close-weight version MiniMax is serving and it is *incredible* in a 24/7 agentic loop. I say this as a self-proclaimed M2.5 hater - *keep your eyes on M2.7*
Not only that but I imagine it sucks from minimax's perspective because they put all this work into a model only for other providers to eat their lunch. If it means more "open weights" models, I don't mind the terms of the license even if they're more restrictive (I mean if it results in more models being freely available and not going closed source).
Absolutely. We're absolutely not blaming anyone. On the contrary. This type of post is meant to show how much we appreciate and patiently await each release.
That's less than 2 hours away! I hope the Unsloth brothers got early access and will have their quants ready at the same time.
Yo la uso con claude code. Hoy entregué para pruebas un proyecto completo que logre hacer con ella en dos días, muy buena. Eso sí, si no sabes lo que le estas pidiendo o como lo está haciendo, estas perdido, eso aplica con cualquier modelo.
Zero-day support and all we get is wah-wah. It blows my mind how pathetically intolerant people have become with open source developers valuable time. Think of this as a free driver update, and be grateful rather than having a sook. The unsloth lads have had every chance to take multi-million dollar jobs at frontier labs and instead they support us, and yet still have yo put up with this lazy whinging about free downloads. Pull your head in.
This is never targeting single users, small companies, ones who generated code. It is targeting infra providers, who would earn money from their work without paying royalties.
You can use it for work. Your work just can't be hosting the model and charging for access (API)
Positive
As if we could stop waiting :) Anything open-source that costs so much money and effort is always worth the wait, and the community is as grateful as ever
From my testing glm, especially glm 5.1, is better in general. But minimax is much smaller and punches well above its weight
M2.7 is the most excited I've been for a release in a while. I've been using the close-weight version MiniMax is serving and it is *incredible* in a 24/7 agentic loop. I say this as a self-proclaimed M2.5 hater - *keep your eyes on M2.7*
Not only that but I imagine it sucks from minimax's perspective because they put all this work into a model only for other providers to eat their lunch. If it means more "open weights" models, I don't mind the terms of the license even if they're more restrictive (I mean if it results in more models being freely available and not going closed source).
Absolutely. We're absolutely not blaming anyone. On the contrary. This type of post is meant to show how much we appreciate and patiently await each release.
That's less than 2 hours away! I hope the Unsloth brothers got early access and will have their quants ready at the same time.
Yo la uso con claude code. Hoy entregué para pruebas un proyecto completo que logre hacer con ella en dos días, muy buena. Eso sí, si no sabes lo que le estas pidiendo o como lo está haciendo, estas perdido, eso aplica con cualquier modelo.
Zero-day support and all we get is wah-wah. It blows my mind how pathetically intolerant people have become with open source developers valuable time. Think of this as a free driver update, and be grateful rather than having a sook. The unsloth lads have had every chance to take multi-million dollar jobs at frontier labs and instead they support us, and yet still have yo put up with this lazy whinging about free downloads. Pull your head in.
This is never targeting single users, small companies, ones who generated code. It is targeting infra providers, who would earn money from their work without paying royalties.
You can use it for work. Your work just can't be hosting the model and charging for access (API)
Negative
why don't they ever compare with their peers. I want to see how it compares to GLM-5.1, KimiK2.5, Qwe3.5-297B, etc.
Casi siempre esta caido porque los servers van sobrecargados
From the looks of it, they want to keep it in a legally grey area so they always have the benefit of doubt. I would avoid this for any serious work tbh.
the minimax hype cycle is getting kinda ridiculous at this point. announce on twitter, post weights on hf "coming soon", then radio silence for 2 weeks. at least when meta drops something they just drop it
licence is really bad lol we won’t even get third party providers so once minimax stops hosting it’ll be gone via api for a lot of people
It's an awful practice. They named the license "modified-MIT" while stripping out commercial use - the core of what MIT is. This kills the open source spirit, and if nobody says anything, every other model provider will do the same thing. It becomes the norm. That's why you call it out.
This is lawyer-speak. Notice this soundbite doesn't mention using your 'code writing' commercially, but his earlier tweet does (it's not allowed).
It is about Unsloth. I don't even know what I'm supposed to say. They just release broken quants all the time and people should not get them.
That's pretty bad, I thought they would verify the quants before uploading, but that would explain why they are always so fast. Bartowski takes longer, probably because he verifies them. I have been using Q4\_K\_XL for a while but with Qwen 3.5 I got curious and compared the quant the q4\_k\_m. Found out q4\_k\_m had better quality in my tests so now I'm back to Bart.
Still NC. Still "modified" MIT. Bruh, just call it NC and be done with it. (cohere, mistral, etc. use those licenses and it's fine.)
SourceAIB PulseLast analyzed