MiniMax M3 on Run BiOS

MiniMax M3 is MiniMax's frontier open-weight model: a native multimodal mixture-of-experts model with a 1M-token context. It has ~428B total parameters with ~23B activated per token. M3 reaches frontier-level performance on specialized tasks such as coding and agentic work, supports image and video input, and can operate a desktop computer. MiniMax describes these three capabilities — frontier coding, million-token context, and native multimodality — as table stakes for closed-source frontier models, and M3 as the first open-weight model to bring all three together.

Highlights

• Native Multimodality — M3 undergoes mixed-modality training from the very first step, enabling deeper semantic fusion across text, image, and video. The entire text pretraining data pipeline was re-architected for the M3 cycle, producing a large volume of interleaved data; MiniMax's experiments show interleaved data scales more easily than synthetic data.
• Context Scaling via Sparse Attention — M3 introduces MiniMax Sparse Attention (MSA), a new sparse attention architecture designed for million-token contexts. Compared with GQA, MSA dramatically reduces attention compute and memory footprint while preserving model quality; across MiniMax's ablations, MSA matched full attention on the vast majority of capabilities.
• Coding & Cowork Capability — M3 achieves frontier-level performance across long-horizon agentic benchmarks, excelling in both coding and cowork.

Architecture: MiniMax Sparse Attention (MSA)

Context scaling was one of the biggest challenges in training M3, and MiniMax addressed it at the attention mechanism itself, avoiding full attention's quadratic computational growth. MSA partitions the KV cache into blocks more precisely than approaches like DSA and MoBA, achieving higher effective context coverage. At the operator level, a "KV outer gather Q" approach uses KV blocks as the outer loop to aggregate the queries that hit them — each block is read only once and memory access is contiguous; under M3's head configuration, MiniMax measures this as more than 4× faster than the open-source Flash-Sparse-Attention and flash-moba operators.

The practical results, as published by MiniMax: at a context length of 1 million tokens, M3's per-token compute is just 1/20 that of the previous-generation model (M2), with a speedup of more than 9× in the prefilling stage and more than 15× in the decoding stage.

Benchmark results

Headline results published in the MiniMax M3 release announcement:

Benchmark	MiniMax M3
SWE-Bench Pro	59.0%
Terminal-Bench 2.1	66.0%
SWE-fficiency	34.8%
KernelBench Hard	28.8%
MCP Atlas	74.2%
BrowseComp	83.5
OSWorld-Verified	70.06% (Max Steps = 200; 68.70% at 100)
PostTrainBench	0.37

As reported by MiniMax: on BrowseComp, M3's 83.5 surpasses Claude Opus 4.7's 79.3, demonstrating strong autonomous browsing and information retrieval. On PostTrainBench — where the model must autonomously complete data synthesis, training, evaluation, and iteration over four base models within 12 hours — M3 scored 0.37, slightly below Claude Opus 4.7 (0.42) and GPT-5.5 (0.39) but clearly ahead of the other models tested.

Evaluation methodology (as published by MiniMax)

• SWE-Bench Verified / Pro: tested on MiniMax's internal infrastructure using Claude Code as the scaffolding (default system prompt overridden for Verified); each test run 4 times and averaged; Pro testing logic aligned with the official evaluation.
• Terminal-Bench 2.1: internal infrastructure, sandbox 8C16G, 2-hour timeout, max output 128K tokens, Terminus 2 scaffolding; scores for GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.7 are taken from the official Terminal-Bench 2.1 leaderboard, all other models tested via official API on the same infrastructure.
• SWE-fficiency: the open-source SWE-fficiency dataset and workflow, sandbox 1C2G, 2-hour timeout, Claude Code scaffolding.
• KernelBench Hard: evaluated with Claude Code on NVIDIA Blackwell GPUs (sm_120); per-question score is the agent's submitted operator TFLOPs relative to the hardware's theoretical peak; benchmark score is the average of 9 questions.
• MCP Atlas: the official MCP Atlas codebase; Public Set scores use Gemini 2.5 Pro as the scoring model, aligned with the official model.
• BrowseComp: the same agent framework as WebExplorer (Liu et al., 2025); when token usage exceeds 64K, all history is discarded.
• OSWorld-Verified: 361 samples from the nogdrive collection using the official OSWorld-Verified codebase; M3 uses relative coordinates 0–1000, 1920×1080 images, Max Steps = 200.
• PostTrainBench: Claude Code with the Ralph-Loop mechanism for 12 hours, testing 4 base models across 5 benchmarks that do not require an LLM judge (AIME2025, BFCL, GPQA Main, GSM8K, HumanEval).
• Video understanding notes: Video-MME at 1 FPS with subtitles interleaved every 30 seconds (M3 scored 84.6 at 512 frames); VideoMMMU at 1 FPS, max 512 frames, LLM-as-a-judge scoring. M3 evaluations used max output 16K–32K tokens with temperature 1.0 and top_p 0.95.
• IMO 2025 & USAMO 2026: aligned with the MathArena official evaluation; dual strong-model grading with human expert rubrics, taking the minimum of the dual judges; M3 evaluated with 512k max output tokens and a test-time-scaling framework of up to 10 iterations.

Real-world tasks (from MiniMax's internal testing)

• Independent paper reproduction — Given an ICLR 2025 Outstanding Paper (Learning Dynamics of LLM Finetuning), M3 ran autonomously for nearly 12 hours, independently producing 18 commits and 23 experimental figures, and successfully completed the core experiments: it matched the trend of prediction-probability changes during the SFT stage, observed the squeezing effect highlighted in the DPO experiments, and verified the Extend mitigation method proposed in the paper.
• CUDA kernel optimization — Asked to optimize an FP8 GEMM kernel for NVIDIA Hopper GPUs starting from only a task description, a benchmark script, and a non-running Triton skeleton, M3 executed continuously for approximately 24 hours, completed 147 benchmark submissions and 1,959 tool calls, and improved Hopper FP8 hardware peak utilization from 7.6% in its first version to 71.3% — a 9.4× speedup — across six landmark rounds of optimization. Its best solution appeared on the 145th submission, after multiple performance plateaus; MiniMax notes that except for Claude Opus 4.7 and M3, most other models stopped making new progress within the first 30 submissions.

Reasoning modes

M3 supports three reasoning modes through its thinking parameter:

• enabled — reasoning is always enabled; suited to complex reasoning, agentic tasks, and long-horizon collaboration.
• adaptive — M3 automatically determines when additional reasoning is beneficial.
• disabled — reasoning is disabled to minimize latency and maximize throughput; suited to conversation and code completion.

MiniMax recommends temperature=1.0, top_p=0.95 for best performance.

On Run BiOS

On the Run BiOS platform, MiniMax M3 is available as the serverless model minimax-m3 with a 1,048,576-token context window, function calling, adaptive reasoning, and prompt caching for repeated prefixes.