Overview
MiniMax M3 is a 428B-parameter mixture-of-experts model (23B active) released June 1, 2026, marking a significant leap for MiniMax's M-series. Its headline innovation is MiniMax Sparse Attention (MSA), a two-stage attention mechanism that pre-filters key-value blocks before running full attention on selected regions, enabling practical 1M-token context at 1/20th the per-token compute of its predecessor. M3 is positioned as the first open-weights model to simultaneously offer frontier-tier coding performance, million-token context, and native multimodal input (text, image, video).
On benchmarks, M3 sits in a competitive middle ground: it slightly edges GPT-5.5 on SWE-Bench Pro (59.0% vs 58.6%), achieves 93% on GPQA Diamond, and scores 83.5% on BrowseComp (surpassing Claude Opus 4.7's 79.3%). It reaches 80.5% on SWE-Bench Verified and 55 on the Artificial Analysis Intelligence Index, placing it level with other leading open-weights models like Kimi K2.6. However, it trails Claude Opus 4.7 on most system-level and agentic benchmarks, and shows weaker performance on abstract reasoning tasks.
The commercial strategy is aggressive pricing: API costs are $0.30/$1.20 per million input/output tokens (standard, ≤512K context), roughly 1/10th to 1/20th of closed frontier models. Subscription Token Plans range from $20–$120/month with generous token quotas. The model is available via MiniMax API, OpenRouter, and planned self-hosting once weights release on Hugging Face (promised within ~10 days of launch). For developers, M3 represents a new cost-performance tradeoff: not the absolute best model, but potentially the best value for coding, long-context, and agentic workloads.
Benchmarks & Performance
MiniMax M3's benchmark performance spans coding, agentic, multimodal, and long-context domains. Below is a detailed comparison with key competitors (all scores from MiniMax's official report unless noted otherwise, with third-party verification from Artificial Analysis and LLMReference):
### Coding & Software Engineering
| Benchmark | MiniMax M3 | Claude Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-Bench Verified | 80.5% | 87.6% | ~84% | ~78% |
| SWE-Bench Pro | 59.0% | 64.3% | 58.6% | 54.2% |
| Terminal-Bench 2.1 | 66.0% | 66.1% | 78.2% | 70.0% |
| KernelBench Hard | 28.8% | 30.7% | 20.9% | 18.6% |
| SWE-fficiency | 34.8% | — | — | — |
### Agentic & Browsing
| Benchmark | MiniMax M3 | Claude Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| BrowseComp | 83.5% | 79.3% | 84.4% | 85.9% |
| MCP Atlas | 74.2% | 77.0% | 75.3% | 69.2% |
| OSWorld-Verified | 70.0% | 82.8% | 78.7% | 76.2% |
| PostTrainBench | 0.37 | 0.42 | 0.39 | — |
### Multimodal & Reasoning
| Benchmark | MiniMax M3 | Claude Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| GPQA Diamond | 93% | ~95% | ~93% | ~90% |
| MMMU Pro | 78.1% | ~80% | 79.9% | ~77% |
| Video-MME v2 | 85.4 | — | — | — |
| GDPval-AA | ~1670 | ~1890 (Opus 4.8) | 1769 | ~1500 |
### General Intelligence
| Metric | MiniMax M3 | Notes |
|---|---|---|
| AA Intelligence Index | 55 | Top open-weights model pending weight release |
| HLE | 37% | Up from M2.7's 28% |
| IFBench | 83% | Up from M2.7's 76% |
| AA-Omniscience | 15.0% accuracy | Only attempts 30.9% of questions (heavy abstention) |
### Efficiency Gains (MSA vs Full Attention at 1M context)
- Per-token compute: 1/20th of M2
- Prefill speedup: >9×
- Decode speedup: >15×
- Compared to open-source Flash-Sparse-Attention: >4× faster
Detailed Comparison
### MiniMax M3 vs Claude Opus 4.7/4.8
Claude Opus remains the strongest closed-source model for system-level reasoning, OS interaction, and high-stakes coding workflows. It scores 87.6% on SWE-Bench Verified (vs M3's 80.5%) and dominates on OSWorld-Verified (82.8% vs 70.0%). However, Opus costs roughly $15/$75 per million tokens—50–60× more expensive than M3's standard pricing. For teams running high-volume agentic loops, M3 offers a dramatically cheaper alternative that captures ~85–90% of Opus's coding capability.
### MiniMax M3 vs GPT-5.5
GPT-5.5 excels at terminal execution (78.2% on Terminal-Bench 2.1 vs M3's 66.0%) and interactive environments. M3 edges GPT-5.5 on SWE-Bench Pro (59.0% vs 58.6%) and leads on SVG generation (SVG-Bench: 63.7% vs 58.2%). GPT-5.5 is stronger on GDPval rubrics (80.6% vs 74.7%). Pricing favors M3 significantly: GPT-5.5 costs approximately $10/$30 per million tokens. The best approach is model routing—use GPT-5.5 for terminal-heavy and strict interactive work, M3 for long-context, repository-scale, and cost-sensitive tasks.
### MiniMax M3 vs Gemini 3.1 Pro / 3.5 Flash
Gemini 3.1 Pro leads on browsing tasks (85.9% BrowseComp vs M3's 83.5%) and has strong multimodal capabilities. Gemini 3.5 Flash scored 84.3% on MMMU Pro, ahead of M3's 78.1%. Gemini's pricing is competitive (~$1.25/$5 per 1M tokens for Pro), but M3's 1M context window with MSA efficiency gives it an edge for truly long-context workloads. Gemini remains the top pick for web design and browsing-centric agent tasks.
### MiniMax M3 vs Kimi K2.6
Both are open-weights models with similar Intelligence Index scores (M3: 55, K2.6: 54). K2.6 scores 79.4% on MMMU Pro, comparable to M3's 78.1%. M3 distinguishes itself with native video input and the MSA architecture enabling practical 1M-token usage. K2.6 is available via Moonshot's platform with different pricing structures.
### Context Window & Pricing Summary
| Model | Context | Input/1M | Output/1M | Open Weights |
|---|---|---|---|---|
| MiniMax M3 | 1M | $0.30 | $1.20 | Pending (~10 days) |
| Claude Opus 4.7 | 200K | ~$15 | ~$75 | No |
| GPT-5.5 | 256K | ~$10 | ~$30 | No |
| Gemini 3.1 Pro | 1M+ | ~$1.25 | ~$5 | No |
| Kimi K2.6 | 128K | ~$0.50 | ~$2.00 | Yes |
Community Feedback
The developer and research community has responded to MiniMax M3 with cautious but genuine enthusiasm. Thomas Wiegold, a full-stack engineer and model reviewer, called it 'one of the more genuinely interesting models I've tested this year' after running it through his standard battery of website builds, poker simulation, and code audit tests. He noted M3 'nailed everything else' on a coffee roaster website prompt—'right up there with the closed frontier models'—and that its code audit was 'remarkably close' to GPT-5.5's quality with 'no filler, no padding, no inventing problems to look busy.' However, he flagged excessive token consumption during complex reasoning tasks (a 30–40 minute poker simulation) and expressed concern about the restrictive license likely to accompany the open weights.
Binary Verse AI positioned M3 as a serious 'systems proposal' rather than just another benchmark release, emphasizing that its sparse attention architecture is 'built for deployment rather than conference applause.' They recommended a 'two-lane setup' combining retrieval layers with M3's expanded workspace for optimal results.
The Stackademic reviewer (TensAI) found M3 'fully competitive' on feature writing and noted it produced the 'most detailed vulnerability report' in an agentic workflow test, though test-writing results were 'the most mixed' among the three models tested. Their verdict: 'a brand-new model matching established leaders on real coding, debugging, refactoring, and agentic tasks is more than a benchmark achievement.'
Artificial Analysis positioned M3 as 'the leading open weights model, once the weights are released,' noting it improved on M2.7 across most evaluations with a +9 point gain on HLE and +6 on GPQA Diamond. The community is closely watching for the weight release and license terms, with many developers adopting a hybrid routing strategy—using M3 for bulk, cost-sensitive, long-context work while reserving closed frontier models for tasks requiring the last few quality percentage points.
On GitHub, the MiniMax-M3 repository has accumulated 409 stars and 45 forks, with active discussions around local deployment via SGLang, vLLM, and KTransformers. The 234,750+ downloads on Hugging Face indicate strong developer interest in self-hosting once fully released.
Use Cases
### 1. Repository-Scale Code Understanding and Agentic Coding
M3's 1M-token context window and strong SWE-Bench scores make it ideal for tasks that require understanding entire codebases at once: repository Q&A, cross-file refactoring, security audits, and multi-file code generation. The CUDA kernel optimization demo (147 iterations, 9.4× speedup, zero human intervention) demonstrates long-horizon autonomous coding capability. Choose M3 over Opus when cost is a significant factor and the task benefits from massive context rather than maximum single-shot precision. Example: giving M3 an entire monorepo and asking it to identify and fix cross-cutting concerns across 50+ files.
### 2. Long Document Analysis and Research Paper Processing
With 1M tokens of practical context (enabled by MSA's efficiency), M3 can ingest entire books, lengthy research papers with figures, or massive log files in a single prompt. The paper reproduction demo (12 hours, 18 commits, 23 figures) showcases this capability. Use M3 when you need to analyze documents exceeding 200K tokens where other models force you to chunk and summarize. Example: loading a full legal contract corpus (500K+ tokens) for due diligence analysis, or reproducing experiments from a complete ICLR paper including code and data.
### 3. Cost-Sensitive Agent Workflows at Scale
For production agent systems running thousands of tasks daily (code review bots, automated testing pipelines, customer support agents), M3's pricing ($0.30/$1.20 per 1M tokens) makes it economically viable at volumes where Opus or GPT-5.5 would be prohibitively expensive. The prompt caching feature ($0.06/1M read) further reduces costs for repeated context. Example: a CI/CD agent that reviews every pull request, runs security scans, and generates test suggestions—processing hundreds of PRs daily at a fraction of frontier model costs.
### 4. Multimodal Document and Video Understanding
M3's native multimodality (text + image + video input, trained from Step 0) enables tasks like processing scanned documents, analyzing technical diagrams, understanding video content, and computer use (operating desktop applications). Scored 85.4 on Video-MME v2 and 78.1% on MMMU Pro. Choose M3 over text-only models when your workflow requires understanding visual inputs alongside text. Example: an agent that opens a local ERP client, reads invoice data from Excel screenshots, and batch-enters information across applications.
Latest News
• **June 1, 2026**: MiniMax M3 officially released as the first open-weights model combining frontier coding, 1M context, and native multimodality. API and Token Plans available immediately.
• **Weight Release Pending**: MiniMax has committed to releasing model weights on Hugging Face and GitHub within ~10 days of launch (expected by ~June 11, 2026). Technical report (arXiv:2606.13392) is already available.
• **Token Plan Updated**: Three subscription tiers launched—Plus ($20/mo, ~1.7B tokens), Max ($50/mo, ~5.1B tokens), Ultra ($120/mo, ~9.8B tokens)—with all three tiers fully available. Text, image, speech, and music share the same usage pool.
• **API Pricing Structure**: Standard rate ($0.30/$1.20 per 1M input/output) for ≤512K context; higher rate ($0.60/$2.40) for 512K–1M context. Thinking mode can be toggled on/off at request time with same pricing. Priority service tier available through sales support.
• **MiniMax Code Updated**: New agent product built on M3 with Agent Team capability (multi-stage, concurrent workflows), Producer + Verifier adversarial loop, and computer use support. Built on open-source OpenCode and Pi frameworks; planned open-source release.
• **Model Availability**: Accessible via MiniMax first-party API, OpenRouter, SiliconFlow, GMI, and Novita. Local deployment supported through SGLang, vLLM, Transformers, KTransformers, and unsloth.
• **MSA Technical Report**: arXiv:2606.13392 published June 11, 2026, detailing the MiniMax Sparse Attention architecture.
• **HuggingFace Activity**: Repository has 234,750+ downloads, 409 GitHub stars, and 31 Spaces using the model within the first month of release.