MiniMax M3: First Open-Weight AI Model With 1M Context
Quick Answer: MiniMax M3 is the first open-weight AI model combining frontier coding, a 1-million-token context window, and native multimodal support (text, images, video). Released June 1, 2026, it scores 59% on SWE-Bench Pro – beating GPT-5.5 – at a fraction of the cost. Free weights arrive on Hugging Face by mid-June 2026.
Open-source AI has spent the past two years chasing a single benchmark at a time: Llama 4 brought strong reasoning, Mistral Small 4 brought compact efficiency, DeepSeek V4 brought scale. MiniMax M3 is the first model to bring all three frontier capabilities – frontier-level coding, a genuine 1-million-token context window, and native image and video understanding – under a single open-weight release. For developers who have been stitching together multiple specialized models for different parts of the stack, M3 is a meaningful consolidation.
According to MiniMax’s official launch post, M3 reaches frontier performance on SWE-Bench Pro (59.0%), Terminal-Bench 2.1 (66.0%), and MCP Atlas (74.2%), while beating Claude Opus 4.8 on SVG generation and BrowseComp. The API is live as of launch. The weights and a full technical report are promised on Hugging Face and GitHub within 10 days – putting the expected open-weight release around June 10-11, 2026.
What is MiniMax M3 and why does it matter for open-source AI?
MiniMax M3 is a multimodal large language model developed by MiniMax, a Shanghai-based AI lab. It launched on June 1, 2026 and is currently available via API. The open-weight release – where anyone can download and self-host the model – is scheduled within 10 days of launch.
What sets M3 apart from existing open-weight releases is the combination of three capabilities that were previously only available in closed-source frontier models:
Frontier coding and agentic work – M3 scores 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, and 74.2% on MCP Atlas, placing it at or above closed-source models like GPT-5.5 on these tasks.
1-million-token context window – The full 1M context is supported through the new MSA architecture. In practice, MiniMax guarantees a minimum 512K context, with 1M available at 2x the standard token rate.
Native multimodal input – M3 natively processes text, images, and video without requiring a separate vision model in the pipeline. It can also operate a desktop computer autonomously, which opens up computer-use agentic applications.
For context on why this combination is significant, compare it to the Llama 4 Scout and Maverick releases, which brought multimodal input but stopped short of frontier coding scores or a genuine 1M context window at speed. M3 is the first open-weight model to check all three boxes simultaneously.
How does MiniMax Sparse Attention (MSA) make 1M context practical?
Standard transformer models use full attention, which has quadratic computational complexity: doubling the context length quadruples the compute cost. At 1 million tokens, full attention becomes economically and practically untenable for most inference deployments. This is why most models with advertised long contexts use them only for very specific use cases and at high cost.
MiniMax Sparse Attention (MSA) addresses this by partitioning the key-value cache into blocks and adding a precision pre-filtering stage before attention. Rather than each query attending to every key, MSA identifies which KV blocks are relevant and reads only those. The system uses a “KV outer gather Q” approach where KV blocks form the outer loop and queries are aggregated against each block – this means each block is read once, memory access stays contiguous, and arithmetic intensity stays high.
The results are measurable. According to MiniMax’s benchmarks, compared to their previous M2 generation at 1M context:
Decoding speed: 15.6x faster (compared to M2 at 1M context)
Prefill speed: 9.7x faster (compared to M2 at 1M context)
Per-token compute: 1/20th of the previous generation at 1M context length
Flash-Sparse comparison: More than 4x faster than open-source Flash-Sparse-Attention implementations
Crucially, across ablation studies MiniMax reports that MSA matches full attention quality on the vast majority of capabilities – the speed gains don’t come at the cost of model quality. This is the architectural claim that makes M3’s long-context performance believable rather than just a marketing number. As APIDog’s technical breakdown notes, MSA’s “clean, scalable, easy-to-implement, and hardware-friendly characteristics” mean the theoretical gains are fully realized in practice rather than being limited to ideal conditions.
How does M3 benchmark against GPT-5.5, Claude Opus, and Gemini?
Benchmark comparisons for frontier models are always noisy, but M3’s scores across multiple independent evaluations form a consistent picture: it is competitive with the best closed-source models on most tasks and leads on several specific benchmarks.
Benchmark MiniMax M3 GPT-5.5 Claude Opus 4.8 Gemini 3.1 Pro What It Measures
SWE-Bench Pro 59.0% 58.6% 69.2% 54.2% Agentic coding (real GitHub issues)
Terminal-Bench 2.1 66.0% 72.1% 74.2% 70.0% Command-line task execution
SVG-Bench 63.7%
59.2%
Visual code generation
BrowseComp 83.5%
–
Autonomous web browsing accuracy
MCP Atlas 74.2%
82.2%
Multi-step tool use
The most significant number is SWE-Bench Pro: M3 edges out GPT-5.5 (59.0% vs 58.6%) while costing roughly 10 to 12 times less. Claude Opus 4.8 still leads by a meaningful 10 points on coding, and on Terminal-Bench all three closed-source models outperform M3. But on SVG generation, BrowseComp, and especially OmniDocBench (multimodal document understanding), M3 holds or leads. As The Decoder reported, M3’s performance-per-dollar ratio is what makes enterprise teams pay attention – for most production agentic workloads, GPT-5.5-level performance at 10% of the cost is a compelling trade.
What does MiniMax M3 cost and how can you access it today?
MiniMax M3 is available immediately via API. There are three access paths, with weights-based self-hosting coming in the next few days.
Input (up to 512K context) $0.60 / M tokens
Output (up to 512K context) $2.40 / M tokens
Cache reads $0.12 / M tokens
Long context (512K-1M) 2x standard rate
Launch discount (7 days) 50% off all tiers
Token subscription plan From $20 / month
The three access options are:
Direct API – Standard token-based billing at the rates above. Two variants are available: M3 (standard) and M3-highspeed (same quality, faster inference). Automatic KV caching is enabled by default, which helps significantly for long-context repeated queries.
MiniMax Code interface – A dedicated coding interface at code.minimax.io, comparable to GitHub Copilot or Claude’s artifacts view, built specifically around M3’s agentic coding capabilities.
OpenRouter – M3 is available via OpenRouter at the model ID minimax/minimax-m3, with the 50% launch discount applied. This gives access to anyone already integrated with OpenRouter’s model comparison infrastructure.
Token subscription plan – MiniMax offers a flat monthly subscription starting at $20 per month for approximately 1.7 billion tokens, scaling up to $120 per month for 9.8 billion tokens. This is significantly cheaper per-token than pay-as-you-go for high-volume workloads.
To put the pricing in perspective: running a 10,000-token prompt through M3’s API costs roughly $0.006, compared to approximately $0.07 for an equivalent GPT-5.5 call. For teams running thousands of agentic coding tasks per day, this difference compounds quickly. This pricing dynamic is part of a broader trend we have been tracking in AI pricing changes for June 2026.
When can you download the MiniMax M3 weights for free self-hosting?
MiniMax confirmed at launch that both the model weights and the full technical report will be published on Hugging Face and GitHub within 10 days of the June 1, 2026 release date. That puts the target window at approximately June 10-11, 2026.
A few important clarifications about what “open-weight” means for M3:
What you will get: The pre-trained model weights in a downloadable format (expected safetensors or similar), plus the technical report detailing the MSA architecture, training approach, and benchmark methodology.
What you will not get: Training data, training code, or RLHF/fine-tuning pipeline details. M3 is open-weight, not fully open-source in the way that, say, Gemma 4 or Mistral models have released training infrastructure.
Hardware requirements: MiniMax has not yet published minimum hardware requirements for self-hosting M3. Given the model’s frontier performance level, expect significant GPU VRAM requirements – likely 40GB+ for full precision, or 24GB+ with quantization. Hardware guidance should arrive with the technical report.
License: The specific license terms have not been confirmed as of this writing. Check the Hugging Face repository when weights drop for commercial use terms, redistribution rights, and any use restrictions.
For developers who want to use M3 immediately without waiting for weights, the API route with automatic caching is the practical path. The MiniMax API’s automatic KV caching is particularly valuable for agentic workloads that share common prompt prefixes across many requests, since cached token reads cost only $0.12 per million – 80% less than input tokens.
This release continues a broader pattern in 2026 where Chinese AI labs are consistently matching or exceeding US frontier model performance on open-weight releases. The combination of benchmark-competitive scores, aggressive pricing, and a clear commitment to weight release makes M3 one of the most significant open-source AI announcements of the year. If you are tracking the free and low-cost AI tier landscape, M3’s pricing structure – especially the subscription plan – changes the math for high-volume agentic deployments that were previously cost-prohibitive with closed-source models.
🔑 Key Takeaways
MiniMax M3 is the first open-weight model to combine frontier-level coding, a 1-million-token context window, and native image and video understanding in a single release, filling a capability gap that previously required multiple specialized models.
On SWE-Bench Pro, M3 scores 59.0% versus GPT-5.5’s 58.6%, while costing roughly 10 to 12 times less per million tokens – a cost-performance ratio that directly challenges closed-source incumbents for production agentic coding workloads.
The MiniMax Sparse Attention (MSA) architecture delivers 15.6x faster decoding and 9.7x faster prefill versus the previous M2 generation at 1M-token context lengths, making long-context inference economically practical rather than a theoretical feature.
Model weights and a full technical report will be published on Hugging Face and GitHub within 10 days of the June 1, 2026 launch, expected around June 10-11, 2026 – allowing self-hosted deployment for teams with sufficient GPU hardware.
API access is live now with a 50% seven-day launch discount, and a subscription plan starts at $20 per month for 1.7 billion tokens, making M3 accessible at scale for teams currently over-paying for equivalent closed-source models.
Related Resources
In-depth reviews of AI tools See how the tools behind the headlines actually perform.
AI tools by profession and use case Find the right tool for what you actually do.
AI scam prevention and alerts Stay safe while exploring new AI tools.
Frequently Asked Questions
What is MiniMax M3?
MiniMax M3 is an open-weight AI model released by MiniMax on June 1, 2026. It is the first open-weight model to combine frontier-level coding ability, a 1-million-token context window, and native multimodal support for text, images, and video in a single model. The API is live now, and downloadable weights are expected on Hugging Face and GitHub by around June 10, 2026.
Is MiniMax M3 open source?
MiniMax M3 is open-weight, not fully open source. The model weights and technical report will be published on Hugging Face and GitHub within 10 days of the June 1 launch. You can download and self-host the weights, but training data and training code are not being released. This is the same open-weight approach used by Llama 4 and Mistral, rather than the fully open approach of Apache 2.0 models like DeepSeek V4.
How does MiniMax M3 compare to GPT-5.5?
MiniMax M3 scores 59.0% on SWE-Bench Pro versus GPT-5.5’s 58.6%, giving M3 a narrow lead on agentic coding. M3 also beats GPT-5.5 on SVG-Bench (63.7%) and BrowseComp (83.5%). The cost difference is the bigger story: M3 costs $0.60 per million input tokens compared to GPT-5.5’s significantly higher rates, making M3 roughly 10 to 12 times cheaper for similar real-world coding workloads.
What is MiniMax Sparse Attention (MSA)?
MiniMax Sparse Attention (MSA) is a new attention architecture that avoids the quadratic computational complexity of standard full attention. It partitions the key-value cache into blocks with a precision pre-filtering stage, then uses a KV outer gather Q approach so each block is read only once with contiguous memory access. The result is 15x faster decoding and 9.7x faster prefill versus M2 at 1M context, while matching full attention quality on most benchmarks.
When will MiniMax M3 weights be available to download?
MiniMax committed to publishing M3 weights and a full technical report on Hugging Face and GitHub within 10 days of the June 1, 2026 launch. The expected release window is June 10-11, 2026. Watch the MiniMax Hugging Face organization and GitHub pages for the announcement. Specific hardware requirements and license terms will be clarified in the technical report when it releases.