June 1, 2026, Shanghai AI CompanyMiniMaxQuietly released its flagship modelM3——A collection1 million tokens super long context,Native multi-modal inputandCutting edge programming capabilitiesAll-in-one open weight model.
core competencies
The positioning of MiniMax M3 is very clear:Specialized models for software engineering tasks.
- 1 million Token context window: M3 adopts a non-compressed Key-Value caching mechanism, and there will be no accuracy attenuation in ultra-long contexts.Claude 3.7 Sonnet is 200K.
- Native multi-modal input: M3 supports text, image and video input, and the output is text.
- MSA sparse attention architecture:Multi-head Sparse Attention is the key to M3's ability to maintain reasoning efficiency under a million token window.
- SWE-Bench Pro 59.0%: On the most challenging real-world programming benchmark, M3 surpassed GPT-5.5 and Gemini 3 Pro, topping all open-weight models.
Programming practical performance
The true value of M3 is reflected inLarge-scale coding taskssuperior.
- Whole warehouse awareness: No more manual chunking or truncating of code.
- Long link debugging: Complex bugs often span multiple files and modules.
- Document + code joint understanding: API documentation, internal Wiki and source code are input at the same time, and the model can be cross-referenced to fully understand the intent of the code.
- Automated test generation: Write meaningful unit tests for existing code - not fill in the blanks with templates, but generate targeted test cases after understanding the behavior of the code.
Feedback from early developers on social platforms has been generally positive, with particular praise for its performance in multi-file refactoring and long-context code reviews.
Cost-effectiveness reshapes the programming model market landscape
The most shocking thing about M3 is not only its performance;Breakthrough leadership in cost-effectiveness.
| Model | context window | SWE-Bench Pro | Enter price (per million tokens) | Output price (per million tokens) |
|---|---|---|---|---|
| MiniMax M3 | 1M | ~59% | ~$0.20 | ~$1.10 |
| GPT-5.5 | 128K | ~40% | $75 | $150 |
| Gemini 3 Pro | 1M | ~50% | $1.25-$2.50 | $10-$15 |
| Claude 3.7 Sonnet | 200K | ~53% | $3 | $15 |
The input price of GPT-5.5 is M3375 times, the output price is136 times——The performance of M3 on programming tasks is actually better.
User experience/limitations
advantage:
- Strong programming skills, especially in long context tasks
- Open weight model, self-deployable and fine-tuned, no risk of vendor lock-in
- The price is very competitive and suitable for large-scale and high-frequency automated programming tasks.
- Native multi-modal input, supporting image and video understanding
- MSA architecture performs well in long context reasoning efficiency
Disadvantages and Notes:
- General reasoning and instruction following capabilities are not as good as GPT-5.5 - if you need a general assistant rather than a programming engine, the GPT series is still a more balanced choice
- Chinese support: Although MiniMax is a Chinese company, M3’s training data is mainly based on programming corpus, and Chinese conversation ability is not its core advantage.
- The ecosystem and tool chain are not as mature as OpenAI and Anthropic - API documentation, SDK support, and community resources are still catching up.
- The benchmark results are self-reported data. Third-party independent evaluation has not been completed on a large scale and needs to be verified in actual tasks.
- Multimodal capabilities focus on input understanding and do not support image/video generation
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 8.2 / 10 | Top programming capabilities, multi-modal and long context blessings, but shortcomings in general capabilities |
| 易用性 | 7.5 / 10 | The API is available but the ecological tool chain is not as mature as OpenAI; self-deployment requires a certain technical threshold |
| Cost-effectiveness | 9.5 / 10 | Breakthrough leadership - achieving stronger programming performance at a price of less than 1% of GPT-5.5 |
| 中文支持 | 7.0 / 10 | Programming corpus is the main focus, and Chinese general dialogue is not the core optimization direction. |
| 输出质量 | 8.5 / 10 | The code generation quality is excellent, with no obvious attenuation in long context scenarios, but self-reported benchmarks require third-party verification |
Overall rating: 8.1/10
Applicable scenarios: Automated code review, large-scale code refactoring, long-link bug debugging, test case generation, CI/CD pipeline integration.
Not applicable scenarios: General dialogues that require multiple rounds of complex reasoning, strong Chinese dialogue scenarios, and enterprise-level deployments with strict requirements on supplier ecology and SLA.
Summarize
The release of MiniMax M3 sends a clear signal:The open weight model has equaled or even surpassed the most expensive closed source model in terms of programming capabilities..
For developers, this means that the cost of programming AI tools can be significantly reduced - high-frequency tasks such as code review, automated refactoring, and test generation in CI/CD no longer require high API fees.
Of course, the M3 is no silver bullet.Claude Fable 5 might be more suitable.Efficient, accurate and low-cost programming assistance, the MiniMax M3 deserves serious evaluation. AI Dash——Discover the best AI tools.