> Published on zhichai.net — English translation of a Chinese forum post. The post is dated June 1, 2026, Shanghai.
MiniMax M3 is not framed as "another Chinese GPT clone." The claim: it is the first domestic (Chinese) model to simultaneously reach usable levels on three capabilities that overseas closed-source frontier models all have — Coding, 1M context, and native multimodality — while choosing to open-source. More strikingly, it publicly documented throwing itself into a 24-hour CUDA optimization ordeal, writing an FP8 GEMM kernel from scratch and pushing hardware utilization from 7.6% to 71.3%.
Key points
1. Three "tickets" to the frontier
- Coding: agent-level programming (understanding requirements, using terminals, tool calls, multi-turn iteration), measured by SWE-Bench Pro, Terminal Bench, MCP Atlas.
- Long context: 1M tokens (~1.5M words), required for sustained agent workflows.
- Native multimodality: trained on text, image, and video from step 0, not bolted-on vision.
- Attention is O(n²); at 1M tokens full attention costs 64x that of 128K.
- Existing approaches: DSA (fixed-interval key sampling), MoBA (block selection), FlashAttention (IO-efficient but still quadratic).
- MSA inverts the flow: KV blocks as the outer loop, gathering the queries that hit them ("outer gather Q"). KV blocks are read contiguously once and serve multiple queries — cache-friendly and high compute-to-memory ratio.
- MiniMax reports MSA is 4x+ faster than Flash-Sparse-Attention and flash-moba under M3's head configuration, with capability parity to full attention in ablations.
- At 1M context: 1/20 per-token compute, 9x prefilling speedup, 15x decoding speedup (end-to-end measured).
- Interleaved data (text-image alternating in-sequence) proved more impactful than paired data — cross-modal understanding emerges from sequence-level interleaving, not additive fusion.
- Data pipeline rebuilt to target ~100 trillion training tokens.
- Starting materials: a task description, a benchmark script, a non-runnable Triton skeleton, no reference implementation.
- 24 hours: 147 benchmark submissions, 1959 tool calls, fully autonomous.
- Six optimization stages: baseline (7.6% utilization) → autotune config generation → bottleneck diagnosis → CUDA Graph integration → persistent kernel rewrite → host-side scheduling.
- Result: peak utilization 7.6% → 71.3%, 9.4x speedup.
- Notably: competing models typically gave up within ~30 submissions; M3's best solution came at submission #145, persisting through multiple plateaus — attributed to long-context attention allocation.
- How close is "close to Opus 4.7" on SWE-Bench Pro, exactly?
- The 7.6% baseline in the CUDA task may be a deliberately weak starting point — "from broken to usable" differs from "from good to better."
- Is flash-moba a fair comparison baseline? Full MSA technical details are pending.
- Will the 10-day open-source promise be honored?
The post argues previous Chinese models each had at most two of the three; M3 is the first to hold all three.
2. MSA: MiniMax Sparse Attention
3. Native multimodality
Three stress tests
2.1 Paper reproduction (12 hours, 18 commits, 23 figures)
Given an ICLR 2025 Outstanding Paper ("Learning Dynamics of LLM Finetuning"), M3 autonomously reproduced it: ~12 hours runtime, 18 git commits, 23 experimental figures, matching SFT-stage predicted-probability trends, observing the DPO "squeezing" effect, and validating the Extend mitigation.2.2 CUDA kernel optimization (24 hours, 147 submissions, 9.4x speedup)
2.3 Self-directed model training (12 hours)
Given four pre-trained-only Base models, M3 autonomously handled data synthesis, training, evaluation, and iteration (AIME2025, BFCL, GPQA Main, GSM8K, HumanEval). Final score 0.37, below Opus 4.7 (0.42) and GPT-5.5 (0.39), clearly ahead of others.Benchmark table
| Benchmark | M3 | Comparison | |------|--------|--------| | SWE-Bench Pro | 59.0% | above GPT-5.5, Gemini 3.1 Pro; near Opus 4.7 | | Terminal Bench 2.1 | 66.0% | — | | SWE-fficiency | 34.8% | — | | KernelBench Hard | 28.8% | — | | MCP Atlas | 74.2% | — | | SVG-Bench | — | above Opus 4.7 | | OmniDocBench | — | above Gemini 3.1 Pro | | Claw-Eval | — | highest score | | PostTrainBench | 0.37 | below Opus 4.7 (0.42), GPT-5.5 (0.39) |Product, pricing, open-source
MiniMax Code (agent product, built on OpenCode and Pi Agent): Agent Team workflows, a Producer + Verifier adversarial harness, multi-day autonomous runs, and Computer Use triggered by phone voice commands. Compared to Claude Code's JS-based Dynamic Workflows, MiniMax Code emphasizes deep reflection and continuous correction.
Pricing: Plus ¥49/mo (600M tokens), Max ¥119/mo (1.8B), Ultra ¥469/mo (5.5B) — roughly 15x Claude's token volume at comparable prices. Caveat: token count ≠ model quality; thinking-mode latency and API stability aren't visible in pricing.
Open-source: MiniMax promised to release the technical report and open model weights within 10 days.
Open questions raised by the post
Conclusion
The post argues M3's real significance is a complete capability stack for "engineer-type AI": understanding existing knowledge (paper reproduction), solving known problems (kernel optimization), and exploring open problems (training models). It trails Opus 4.7 and GPT-5.5 on the hardest tier, but the gap is narrowing — and if the open weights ship on time, M3 could become key infrastructure for developers studying agentic engineering AI.References cited in the post
1. MiniMax Research (2026). MiniMax M3: Frontier Coding, 1M Context, Native Multimodal. minimaxi.com/blog. 2. Vaswani et al. (2017). Attention Is All You Need. NeurIPS. 3. Gu & Dao (2024). Mamba: Linear-Time Sequence Modeling with Selective State Spaces. ICML. 4. Dao et al. (2022). FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. NeurIPS. 5. Garg et al. (2022). What Can Transformers Learn In-Context? A Case Study of Simple Function Classes. NeurIPS.*Note: this is a translation of a community post; claims reflect the original author's reporting of MiniMax's announcements, not independent verification by zhichai.net.*