MiniMax M3: Coding Power, 1M Context, and Native Multimodality in One Open Model

> On June 1, 2026, MiniMax released M3 — not just a bigger model, but a complete toolkit: coding ability that reads million-line codebases, 1M-token context, native multimodality, and agent autonomy that can work for hours unattended. Open weights promised within 10 days.

Key points

• M3 combines three capabilities in one open model: frontier coding, 1M-token context, and native multimodal training (text, image, video interleaved from step 0, not bolted on).
• MSA (MiniMax Sparse Attention) reduces the O(N²) cost of long-context attention via block-level routing plus in-block full attention, with an "outer gather Q" design for contiguous memory access. MiniMax reports it is over 4x faster than Flash-Sparse-Attention and flash-moba.
• Benchmarks: SWE-Bench Pro 59.0%, Terminal Bench 2.1 66.0%, SWE-fficiency 34.8%, KernelBench Hard 28.8%, MCP Atlas 74.2%. M3 surpasses GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro and approaches Claude Opus 4.7 (independently verified numbers pending).
• Multimodal evals: tops SVG-Bench (vs Opus 4.7), OmniDocBench (vs Gemini 3.1 Pro), and the Claw-End-to-end agent evaluation.

Company trajectory

• Late 2021: founded in Shanghai by ex-SenseTime engineers; first model family named abab.
• Jan 2025: MiniMax-01 open-sourced with Lightning Attention (linear-complexity long sequences).
• Jun 2025: M1 — 456B total / 45.9B activated parameters, 1M context.
• Oct 2025: M2 — Sparse MoE, 230B total / 10B activated; inference cost ~8% of comparable closed models.
• Feb 2026: M2.5 — 80.2% on SWE-Bench Verified.
• Mar 2026: M2.7 — "recursive self-evolution" pipeline where models help generate their own training data, compressing release cycles to ~1 month.
• Jun 1, 2026: M3 launches; technical report and open weights promised within 10 days (M1–M2.7 were all released under MIT).

Three autonomous-agent case studies

1. Paper reproduction (~12 hours, unattended): M3 reproduced an ICLR 2025 Outstanding Paper on LLM finetuning learning dynamics, producing 18 commits and 23 experiment charts. It used multimodality to read figures, 1M context to hold paper + code + logs, coding to run experiments, and agent loops to iterate — verifying the DPO squeezing effect and the Extend mitigation. 2. FP8 GEMM CUDA kernel (24 hours): starting from only a task description, a benchmark script, and a broken Triton skeleton — no reference implementation — M3 made 147 benchmark submissions and 1,959 tool calls, progressing through autotuning, bottleneck diagnosis, CUDA Graph integration, persistent-kernel rewrite, and host-side scheduling. FP8 utilization rose from 7.6% to 71.3% (9.4x speedup). Most other models gave up within ~30 submissions; M3's best solution came at submission 145. 3. Autonomous post-training (PostTrainBench): given four base models, M3 autonomously ran data synthesis → training → evaluation → iteration within 12 hours across AIME2025, BFCL, GPQA, GSM8K, and HumanEval. Score: 0.37, below Opus 4.7 (0.42) and GPT-5.5 (0.39) but clearly ahead of other models.

Products and pricing

• MiniMax Code: an agentic coding product rivaling Claude Code. Key differences: deep reflection with continuous error correction, multi-stage concurrent Agent Team workflows, native multimodal Computer Use (desktop operation), a Producer + Verifier adversarial harness that can run autonomously for days, and a planned open-source harness.
• Token Plan pricing (roughly 15x Claude's token volume at equivalent prices):
• Plus: ¥49/month, 600M tokens (~5x Claude Pro $20)
• Max: ¥119/month, 1.8B tokens (~2x Claude Max 5x $100)
• Ultra: ¥469/month, 5.5B tokens (~3x Claude Max 20x $200)
• API: open now, two context-length pricing tiers, thinking / non-thinking dual modes, and a priority service tier (service_tier=priority) for high-concurrency SLA-sensitive workloads (currently sales-gated). Existing user pricing is preserved.

Limitations and caveats

• Benchmark context: many M3 results were run on MiniMax's own infrastructure using agent scaffolds (Claude Code, Mini-SWE-Agent, Terminus), conflating model and harness capability. Independent third-party verification is pending.
• Long-context attention quality: MSA reportedly matches full attention on "most" tasks at 1M context, but detailed degradation curves (e.g., needle-in-haystack precision) were not disclosed.
• Seeing ≠ understanding: native multimodality solves format compatibility; whether M3 grasps statistical assumptions behind charts remains to be validated.
• Ecosystem fit: whether MSA adapts smoothly to vLLM, SGLang, and other inference frameworks after open-sourcing needs community testing.

Takeaway

M3's story is not "surpassing GPT" but completing a puzzle: the first domestic model to combine coding, 1M context, and native multimodality — open-source. The case studies point to where coding agents are heading: not single-shot code generation accuracy, but long-horizon autonomy, planning, persistence through plateaus, and human-agent collaboration.

> "Model updates are moving so fast today that it's easy to forget this remains a matter of steady, daily progress. It has its own objective laws, and it rewards teams that advance solidly along them." > — from the closing of MiniMax's M3 launch post

References

1. MiniMax M3 official launch post (2026-06-01). https://www.minimaxi.com/blog/minimax-m3 2. Liu et al. (2025). *MoBA: Mixture of Block Attention for Long-Context LLMs*. Moonshot AI. 3. DeepSeek-AI (2025). *Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention*. 4. Jimenez et al. (2024). *SWE-bench: Can Language Models Resolve Real-World GitHub Issues?* ICLR 2024. 5. Ouyang et al. (2025). *KernelBench: A Benchmark for Automatic Kernel Optimization*.