MiniMax M3 Launches: 59% SWE-Bench Pro, 70% Computer Use, 15x Faster Long-Context Than M2
MiniMax released M3 on June 1, its first model with a native sparse attention architecture that fundamentally changes how it handles long context. The headline numbers: 59.0% on SWE-Bench Pro (ahead of GPT-5.5 and Gemini 3.1 Pro), 70.06% on OSWorld-Verified for computer use, and a 15x decoding speedup at 1M tokens compared to M2.
The architecture change is MSA — MiniMax Sparse Attention — which replaces standard dense attention in the context window. At 1M tokens, MSA delivers more than 9x faster prefill and more than 15x faster decoding at 1/20th the per-token compute versus M2. For agentic workflows that hit the million-token scale, that is a material operational difference, not a benchmark artefact.
Key Numbers
- SWE-Bench Pro: 59.0% (surpasses GPT-5.5 and Gemini 3.1 Pro; approaches Opus 4.7)
- OSWorld-Verified: 70.06% task completion on computer use (Max Steps = 200, 361 samples)
- Terminal-Bench 2.1: 66.0%
- MCP Atlas: 74.2%
- SVG-Bench: surpasses Opus 4.7
- OmniDocBench: above Gemini 3.1 Pro on multimodal document understanding
- Context: 1M tokens (native, MSA-powered)
- SWE-Bench Pro testing: Claude Code scaffolding, averaged over 4 runs, aligned to official evaluation
SWE-Bench Verified numbers were not published at launch — the SWE-Bench Pro figure uses Claude Code scaffolding on internal infrastructure, which requires a caveat. It tracks the same task distribution as the external leaderboard but is not an independent third-party run.
Architecture: MSA Replaces Dense Attention
MSA is a sparse attention scheme built for long-context production workloads. Standard transformers require quadratic compute in sequence length — at 1M tokens that becomes impractical at production scale even with FlashAttention. MSA restructures the computation to avoid attending to every token pair, preserving retrieval accuracy while reducing FLOPs by a factor of 20 versus M2 at full context.
The model is also natively multimodal from the ground up: image and video input alongside text, with desktop computer operation support baked into the architecture rather than patched in post-training. Thinking mode is toggleable at request time, giving developers the option to trade latency for reasoning depth on a per-call basis.
Pricing and Availability
The MiniMax API is live. Token Plan starts at $20/month, covering approximately 1.7 billion M3 tokens. Open model weights and the full technical report are committed for release within 10 days of the June 1 launch.
M3 is the successor to M2.7, which MiniMax open-sourced at 1495 GDPval-AA ELO and 57% Terminal-Bench. M3 pushes both figures up and adds computer use capability at a scale no previous open MiniMax model reached.
Competitive Position
The SWE-Bench Pro result puts M3 in a three-way bracket just below Opus 4.7. GPT-5.5 and Gemini 3.1 Pro both sit below M3 on this benchmark. That is a credible claim for a model from a lab outside the US-China frontier tier — if the numbers hold up under independent evaluation once weights ship.
The OSWorld score is the more surprising figure. Computer use at 70% places M3 above most proprietary alternatives on that specific benchmark. The task distribution in OSWorld-Verified covers real desktop applications, not synthetic environments, which makes it a harder and more meaningful signal than abstract coding tasks.
For teams evaluating agentic stacks, M3 is worth tracking when weights drop. The combination of 1M context at production speed and competitive coding and computer-use scores is a configuration that does not currently exist in the open-weights market at this quality level.