MiniMax M2.7

MiniMaxreleased 的2290亿参数MoE架构代码 and collaborate 专用large model，以Claude Opus约7%的成本完成90%以上代码任务质量

large language model

https://www.minimaxi.com/models/text/m27

Visit Website

In-depth Report

MiniMax M2.7 is the flagship large language model released by MiniMax (Xiyu Technology) on March 18, 2026. It adopts the MoE hybrid expert architecture, with a total parameter volume of 229 billion and approximately 10 billion activation parameters. As the first model in China that is deeply involved in its own training iterations, M2.7's biggest technological breakthrough is the introduction of reinforcement learning Harness automation cycles into the training process, allowing the model to take on about 30% to 50% of the research and development workload. It achieved a score of 56.22% in the software engineering benchmark SWE-Pro test, and its code and bug detection capabilities are close to the level of Claude Opus 4.6. In terms of pricing, the input price of M2.7 is only about 1/15 of Claude Opus 4.6, and is regarded as the most cost-effective quasi-first-tier code model in the first half of 2026. It is worth noting that although the weights of this model have been open sourced, it adopts the MiniMax-Modified non-commercial license agreement, which prohibits unauthorized commercial deployment.
Founded in 2021, MiniMax is a technology company focusing on artificial intelligence technology and is headquartered in China. The founding team of MiniMax has deep experience in the field of AI, and the company has received investment from many well-known institutions. M2.7 is the latest iterative version of the MiniMax M series model. The iteration cycle from M2 to M2.7 is about 6 months, and each generation has significantly improved on key benchmark indicators. The positioning of M2.7 is not a general dialogue model, but a code and professional task model focusing on "Cowork Agent" collaboration scenarios. From a technical perspective, M2.7 adopts a sparse mixed expert (MoE) architecture, with each token activating only about 10 billion parameters (accounting for 4.3% of the total parameters). This design allows the model to significantly reduce inference costs while maintaining strong capabilities. On April 12, 2026, MiniMax officially announced the open source weight of M2.7, but used a special agreement called "MiniMax-Modified Non-Commercial License", which allows personal and research use, but commercial use is clearly prohibited. Commercial use requires a written authorization application to MiniMax in advance. This license arrangement has caused some controversy, and is considered by some developers to be suspected of "pseudo-open source".
The biggest technical highlight of M2.7 is its "self-evolution" ability. The training iteration of traditional models relies on a large amount of manual participation, including data screening, experimental design, evaluation and analysis, etc. M2.7 has built a set of automated reinforcement learning Harness, which can drive experiment monitoring, log inspection, code repair and evaluation cycles by itself. The model can take on about 30% to 50% of the research and development workload during the training process. MiniMax internal evaluation data shows that this mechanism brings about a 30% overall performance improvement. The implementation of this mechanism relies on the Agent Teams capability natively supported by M2.7. Multiple AI agents can work together to complete complex multi-stage tasks, and this capability has been internalized into the bottom layer of the model rather than relying on external prompt word engineering.
The M2.7 performs well on several software engineering benchmarks. It scored 56.22% in the SWE-Pro test, close to the level of GPT-5.3-Codex. Actual measurements by Kilo Code's independent laboratory show that in tests on three TypeScript code bases, M2.7 and Claude Opus 4.6 were able to find all 6 bugs and all 10 security vulnerabilities, with completely equal performance. However, there is a gap in the thoroughness of the fix: Claude Opus 4.6 generated 41 integration tests (including testing the database and cleaning logic), while M2.7 generated 20 unit tests (direct verification of the schema and processing functions). In online operation and maintenance scenarios, M2.7 can independently complete the entire process of analyzing monitoring alarms, connecting to the database to verify causes, non-blocking hemostasis processing (such as index creation), and submitting PRs. Official data shows that based on M2.7, fault recovery time has been compressed to less than 3 minutes many times.
The capabilities of M2.7 in the professional office field have also been fully demonstrated. It scored 1495~1500 in the GDPval-AA ELO evaluation, ranking fourth in the world. Professional scenarios include financial document analysis, revenue forecast modeling, etc. In complex editing scenarios of the three-piece Office suite (Excel/PPT/Word), M2.5 to M2.7 has significantly improved, supporting multiple rounds of modifications and high-fidelity editing. In terms of tool calling (Function Calling), the Toolathon benchmark accuracy rate reaches 46.3%, which is close to the latest Claude Sonnet 4.6 level in the OpenClaw evaluation.
M2.7 adopts MoE sparse hybrid expert architecture, with a total number of parameters of 229 billion and approximately 10 billion activation parameters. Multi-head causal self-attention combines rotational position encoding (RoPE) and QK RMSNorm optimization. The context window is 204,800 tokens (about 200,000), and the maximum output length is officially nominally up to 204,800 tokens. Knowledge deadline is March 2026. The model uses plain text input and does not support multi-modal input such as images, audio or video. The API protocol is fully compatible with the OpenAI format and supports common functions such as Function Calling, Streaming, and JSON Mode.
M2.7 provides two API versions. The standard version M2.7 has an input price of $0.30/million tokens and an output price of $1.20/million tokens; the high-speed version M2.7-highspeed has an input price of $0.60/million tokens and an output price of $2.40/million tokens. The automatic context caching function takes effect without configuration, and the reuse price after caching is reduced to approximately US$0.06/million tokens, significantly reducing the cost of long conversations and code base reference scenarios. Compared with other mainstream models, the price advantage of M2.7 is extremely obvious. Taking input prices as an example, Claude Opus 4.6 is $5.00/million tokens (about 17 times that of M2.7), GPT-5.4 is $2.50, Claude Sonnet 4.6 is $3.00, and DeepSeek V3 is only $0.27. According to the actual test case of Kilo Code, the same test task of M2.7 costs US$0.27 and Claude Opus costs US$3.67, which is about 7% of the cost of competing products. For a team processing 10 million tokens per day, the annualized cost of M2.7 is about $30,000, while Claude Opus 4.6 is closer to $900,000. In addition to pay-as-you-go, MiniMax also offers a Token Plan subscription, with annual fees ranging from US$100 (entry) to US$1,500 (super-fast). Users can also access through various channels such as OpenRouter (ranked 4th in usage), HuggingFace (open weight download), agent.minimax.io (free trial quota), etc. The price of M2.7 on OpenRouter is consistent with the official price, and there is no additional price increase.
The performance of M2.7 in multiple mainstream benchmark tests is as follows: MMLU-Pro 78.2 points (compared to Claude Opus 4.6's 80.1), mathematical reasoning MATH-500 score 91.6 (Claude Opus 4.6 is 93.2, a difference of only 1.6 points), and HumanEval+ code test 87.5 points. But the SWE-Bench Verified (a third-party independent verification version) score of 48.3% is still about 7 percentage points behind the Claude Opus 4.6's 55.7%. The long context needle in the haystack test (NIAH 100K) scored 99.1%, and the 1M ultra-long context scored 96.8%, which are the core advantages of M2.7. Artificial Analysis' AA Intelligence Index v4.0 independent measurement gives it a score of 50 (compared to 53 for Claude Opus 4.6), and the LMSYS Chatbot Arena independent verification score is 1447ELO. It should be noted that some of the high-scoring benchmarks (SWE-Pro 56.22%, Terminal Bench 2 57.0%, VIBE-Pro 55.6%) are MiniMax self-tests or internal benchmarks, and there are two confirmed performance regressions: 11 percentage points lower than M2.5 on the τ²-Bench telecom task, and performance on the BridgeBench vibe-coding task is not as good as M2.5. In addition, Artificial Analysis points out that M2.7 generates approximately 55% more total output tokens than M2.5, which to a certain extent increases the benchmark score for reward completeness, but also means higher costs and slower response times in a real production environment.
Since M2.7 was released relatively recently, public large-scale user evaluation data is limited. Judging from the available information, developers and the AI application community have highly recognized the cost-effectiveness of M2.7. The previous generation product M2.5 has occupied the first position in Code mode usage on the Kilo Code platform, with a market share of 37%, surpassing Claude Opus 4.6 and GPT-5.4. This market foundation provides strong endorsement for the promotion of M2.7. Judging from actual test feedback, M2.7 has received many positive reviews for daily development tasks such as code generation, bug detection, and security vulnerability identification. Developers generally recognize its feature of significantly reducing usage costs while maintaining high-quality output. Some users have reported that M2.7 generates longer responses and is not as good as the Claude series models in handling some scenarios that require concise replies. The text-only limitation has also been criticized by some users who require multimodal capabilities.
The release of M2.7 set off a new round of price war in the code model market, significantly lowering the threshold for using top code models. From the competitive landscape, direct competitors of M2.7 include Claude Opus 4.6, GPT-5.4-Codex, Claude Sonnet 4.6, Kimi K2.5 and DeepSeek V3. In terms of code capabilities, M2.7 is close to Claude Sonnet 4.6, behind Claude Opus 4.6 and GPT-5.4; in terms of price, M2.7 is second only to DeepSeek V3's extremely low pricing, which is far lower than other competing products. The "self-evolution" mechanism of M2.7 has triggered a lot of industry discussions. Supporters believe that this represents a new paradigm for large model training, and that the model's ability to autonomously participate in the optimization process will significantly accelerate the iteration of AI capabilities; critics believe that this mechanism is still an automated evaluation cycle running within a strict programmatic framework, and is not truly autonomous evolution, and that marketing promotions are overly anthropomorphic. Others point out that if MiniMax maintains the current rapid iteration pace, it will put pressure on the pricing strategies of Anthropic and OpenAI and push the entire industry toward lower prices.
Controversy over the open source nature is one of the main doubts faced by M2.7. Although MiniMax announced the open weight of M2.7, the license clearly prohibits commercial use. The first line of the LICENSE file on HuggingFace is marked "NON-COMMERCIAL LICENSE". This is essentially different from open source in the usual sense (represented by MIT, Apache and other protocols, which allow unconditional commercial use). MiniMax's promotion of it as "open source" is considered misleading by some developers. The gradual tightening of the M series license from the M2's MIT agreement has also triggered discussions about the integrity of the company's open strategy. Data security risks are a factor that enterprise users need to consider. MiniMax's parent company is headquartered in China and is bound by Article 7 of the National Intelligence Law of the People's Republic of China. Although the global API (api.minimax.io) is operated by a Singaporean entity and the data is isolated from domestic endpoints, compliance review and audit risks are still significant for users in sensitive industries such as finance, healthcare, and government. If an enterprise has strict requirements for data sovereignty, it is recommended to give priority to self-built deployment solutions, but the hardware threshold needs to be evaluated - the FP16 non-quantized version requires about 461 GB of video memory (about 7 H100 80GB), and the 4-bit quantized version also requires 115~130 GB of video memory. The issue of benchmark trust also deserves attention. The highest-scoring benchmark data (SWE-Pro, VIBE-Pro, Terminal Bench, etc.) are all MiniMax self-tests or internal benchmarks. The data independently verified by a third party is relatively conservative. Enterprise users should take this factor into consideration when selecting models. Additionally, the model is in non-U.S. Entity List status, but there is a supply chain risk of future inclusion.
The best usage scenarios of M2.7 include: daily programming assistance and code generation (accounting for 80% of regular tasks, M2.7 is the most cost-effective), analysis and understanding of ultra-long code bases (200,000 contexts and low input prices have obvious advantages), multi-Agent collaboration pipeline construction (supported by native Agent Teams), high-concurrency SaaS back-end services that need to control API costs, and the generation end of the RAG system (long contexts can accommodate more search results). Scenarios where M2.7 is not recommended include: complex enterprise-level code refactoring and long-term architecture design (Claude Opus 4.6 or GPT-5.4 is recommended), capabilities that require multi-modal processing (models are limited to plain text), non-technical content creation and prose writing (Claude series is better at), and commercial product integration that has strict requirements for license compliance. A pragmatic combination usage strategy is to use M2.7 for daily programming Q&A, document analysis and general tasks (about 80% of the total workload), Claude Opus 4.6 or GPT-5.4 for complex engineering problems (about 20%), and DeepSeek V3 for large-scale batch data processing and content generation. This combination can significantly control API costs while maintaining output quality.
MiniMax M2.7 is a quasi-first-tier, cost-effective large model dedicated to code and collaboration. Its core competitiveness lies in completing more than 90% of the code task quality at about 7% of the cost of Claude Opus 4.6. Coupled with the support of native multi-agent collaboration and self-evolution training mechanisms, it will be extremely competitive in the code model market in the first half of 2026. However, controversies such as its non-commercial license restrictions, geopolitical risks to data security, and insufficient independence of benchmark data cannot be ignored. It is suitable for technical teams that are cost-sensitive and focus on coding tasks, but enterprises should fully evaluate compliance requirements and license constraints before production integration.

User Reviews

EugenHauk

—

From the overall experience, M2.7 has exceeded expectations in terms of office work, coding, and compatibility with MaxClaw. It has helped me save money on BI tools and Coding subscriptions. The flaw is still complex bugs, rapid location and large-scale code rewriting error rate, and contradictory prompts in multiple rounds are not handled well enough. As the first echelon of domestic Agent models, closed source and pure text are the two things I regret most.

丁梅

—

When running a complex Agent, 40 Skills can still maintain an instruction compliance rate of 97%. The 32 skills on my computer are basically not messed up. This is better than most models, and it will not interfere with the installation of too many.

Gary_Peterson_X

—

PinchBench ranks fourth, 0.7 percentage points away from Sonnet 4.6, which is already considered the ceiling of domestic products in the lobster ring.

秦心

—

I actually used it to handle an online slow query failure. It took less than an hour from alarm to root cause location. It also knew to use a non-blocking method to first build an index to stop the bleeding and then add MR. This kind of SRE-level judgment is much more valuable than simply writing code. In daily development, it can handle 70% to 80% of the solution design and coding, and the rest has to be checked by others. The volume is large and the management is full, but there is really no guarantee.

FoxFinanceCooper

—

The context of long processes and multiple rounds of dialogue will decay. It will not proactively confirm contradictory requirements with you, and will directly filter your archived topic selections. For complex systems, it is better to design the product first and then throw it to it.

DRussell_20210

—

It’s not expensive at all. The API is 20 times cheaper and faster than Opus, and cached reads are almost free.

Russell.Cook

—

The point of self-evolution is really scary. MiniMax let M2.7 optimize the internal scaffolding by itself. It ran more than 100 iteration cycles with zero manual intervention. It analyzed the failure trajectory, modified the code, ran evaluations, and compared it to decide whether to retain or roll back. The internal evaluation set directly increased by 30%. Kaggle MLE Lite got 9 golds, 5 silvers and 1 bronze in 24 hours, tying with Gemini-3.1.

天涯_1

—

Let it be used to create an iPhone appearance inventory web page. The structure and timeline are correct. However, the image capture failed to be loaded by the wall. Finally, MaxClaw was used to rub the image before inserting it. The image acquisition was still stuck.

烟雨93

—

The open source party has broken the defense. M2.5 is still Apache2.0, and M2.7 is directly closed source. Commercial self-deployment requires written authorization.

PaulRoss_66755

—

After the Token Plan upgrade, the quota increased four to five times and the pricing was changed to RMB. The Plus plan costs 49 yuan per month for 1,500 requests/5 hours, which is very friendly to domestic developers. That is, the 5-hour rolling window will not be rolled over, and the current will be dynamically limited from 15:00 to 17:30 during the peak period, so you have to use it with time.

GraviqyDrop238

—

Let’s complain, you really can’t gamble on mathematics and complex nested instructions. L-Math Hard only scores 15 points, and the loop output collapses badly. When it comes to reasoning, you should use another model.

bluetiger391

—

Code can be typed better than imagined, SWE-Pro 56 is equal to Opus, engineering work is really good.

Bryan.Gonzalez_2020

—

To be honest, heavy users who burn 3,000 dollars a month Skills: M2.7 responds much faster than Claude. The high-speed version provides instant gratification. Agent Team can also handle multi-role collaboration. It can write a 3D theme park web page and start running with just one sentence. However, you still have to change the reading style of the Office three-piece suite, especially the reading style of PPT and the original character design of long articles. Conclusion: In the three dimensions of Agent collaboration, office automation, and cost-effectiveness, the domestic ceiling has been eliminated.

KathleenWilliams_2020

—

According to actual production troubleshooting, M2.7 can really correlate monitoring indicators and deployment timelines to make causal inferences, locate missing index migration files, and know to first use non-blocking index building to stop bleeding and then raise MR, and recover in three minutes. This is not like a model that can only write code.

ETlop

—

It's a pity that the pure text model is multi-modal this year and it does not support images.

Abigail.PowellJr311

—

MaxClaw is awesome. Just install the Skill and just drop the link. If there is a network problem, it will try again on its own. You don’t have to fiddle with the configuration page. It finally feels like “hosting your brain”.

Ethan.Cook_2022

—

I connected MiniMax M2.7 to Claude Code and ran a 4.4G stock data visualization. It installed Pandas, cleaned, built the model, and finally converted it into an interactive web page using Streamlit. I didn't have much involvement in the whole process. The long-term mission context is really stable, and the nearly 10,000-word research report shows no illusions. This is better than many flagships.

PeterButler_2021

—

M2.7 helped me make a SanDisk financial report in Excel. Five sheets were spit out directly. The dark blue color scheme of the investment bank is too cool.

PamelaPrice_2021

—

The price is so good. It only costs 0.3 knives to enter one million tokens. Claude costs 15 knives!

happyrabbit128

—

After connecting to Cursor and running for a few days, M2.7 is really the king of cost performance. Daily CRUD tasks are not inferior to Claude at all. Occasionally if you encounter difficulties, you can switch to Opus and save a lot of money.

DianaWalker_2022

—

Open source is open source, but the license states that commercial use is prohibited. Everyone, this is not open source in the true sense. Don’t be led away by the propaganda.

JenniferEvans36956

—

I tried deploying it. FF16 requires 7 H100s without quantization, and 4090 can’t run at all... I’d better just use the API.

Declan593

—

Kaggle MLE Lite achieved a card winning rate of 66.6%. This result is second only to Opus 4.6 and GPT-5.4. It is already the highest in China, to be honest.

梅花643

—

Finally open source! But after using it, I found that it cannot be used commercially... Fortunately, the API is cheap, otherwise I would be really scolded.

EJohnsonSr

—

It can be used on OpenRouter, and the switching cost is almost zero. Claude SDK can directly change the environment variables and it will be switched. This experience must be praised.

Steven_Anderson369904

—

I tested the bug detection. I found 6 bugs and it found them all. It was a tie with Opus. There is something.

Ethan_Hicks007305

—

The concept of self-evolution is very scary, but to be honest, it is just running an automated loop in a strict RL framework, which is not that magical.

George_Alvarez_X

—

The stability is really much better than that of Zhipu. Zhipu sells out at every turn. MiniMax has not encountered this situation so far.

NathanMyers_738

—

After comparing the packages of four major domestic models, MiniMax has the best overall solution: stability, large quantity, and reasonable price. Kimi was pressed to the ground and rubbed. The problem of abnormally high token consumption was too big. The actual available amount was directly discounted by 50%.

goldenleopard197

—

The measured speed of the high-speed version is about 7-10 tokens/sec, which is not the 100TPS officially claimed, but it is enough.

MiniMax M2.7

In-depth Report

User Reviews

Similar Products