MiniMax open-sources M3.1 Flash: the 428B "Space Bunny" agent model lands on Hugging Face

• Who: MiniMax, the Chinese AI lab behind the MiniMax M3 family
• What: Open-sourced the MiniMax M3.1 Flash model weights, announced October 23, 2026
• Size: 428B total parameters / 23B active per token (MoE); 1M-token context; text, image and video input
• Where: huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash — repo files updated 2026-10-23T03:37:20Z with the full weight shards
• Ecosystem: day-one vLLM, SGLang, Baidu FlagOS, Hugging Face and ModelScope support; chips from Huawei Ascend, NVIDIA, AMD and more
• Context: "Space Bunny" topped OpenRouter’s charts for 10 days before its name was revealed; API launched October 19 at $0.15/$0.60 per 1M tokens

MiniMax has open-sourced the weights of MiniMax M3.1 Flash. ChainCatcher reported the announcement on October 23, and the weights are live on the official MiniMaxAI/MiniMax-M3.1-Flash Hugging Face repository, whose files were updated today to include the full safetensors weight shards. M3.1 Flash is a 428-billion-parameter mixture-of-experts model — 23 billion parameters active per token — built for agentic work: coding, computer use and cybersecurity, with text, image and video input and a 1-million-token context window.

This is the second act of the "Space Bunny" playbook. On October 19, MiniMax confirmed that the anonymous model that had quietly topped OpenRouter’s daily token charts for ten straight days — processing 56.7 trillion tokens on that platform alone, and over 200 trillion in total by MiniMax’s own count — was an early build of M3.1 Flash, as Newscast reported. The name, the $0.15/$0.60 per-million-token pricing, and now the weights arrived only after the model had already won its audience.

Day-one hardware support

The release comes with an unusually broad ecosystem push. MiniMax lists day-one support from Chinese chip makers Huawei Ascend, T-head, Kunlunxin, Moore Threads, Biren, Haiguang, Muxi and Tiandao alongside NVIDIA and AMD; cloud platforms Together AI, Fireworks, Nebius, Baseten, Modal and DigitalOcean; and the inference stacks vLLM, SGLang, Baidu’s FlagOS, Hugging Face and ModelScope. The model card’s recommended deployment is a single hf download command, with SGLang, vLLM and TokenSpeed as the recommended serving frameworks and Transformers, KTransformers, unsloth and ATOM integrations still in progress.

What’s inside

Under the hood it is a native multimodal MoE: sparse attention throughout the model, quantization-aware training, and DSpark speculative decoding — the draft head ships in this release. MiniMax claims roughly 3x the inference speed of M3 and lists about 150 tokens per second on its own API. Five reasoning-effort levels (low to max) stay always on; they cannot be disabled.

The honest read on benchmarks — every one vendor-reported, none independently verified: 62.8 on Automation-Bench v1.0.6 (top of MiniMax’s own table), a 76.2 tie with Anthropic’s Opus 5 on BankerToolBench, and 73.7 on DeepSWE v1.1 — where DeepSeek’s V4.1 Flash scores higher. Terminal Bench 4.0 is the weak spot at 38.1 versus 51.8 for Opus 5. API pricing stays flat at $0.15 input / $0.60 output per million tokens, half of M3’s standard rate. MarkTechPost’s launch rundown has the full comparison table.

The license question

One caveat before you self-host: the Hugging Face repo’s license field is set to "other" — not a standard open license. The files are downloadable; the legal terms are still MiniMax’s to spell out. "Open weights" here means you can run the model yourself, not that the terms are OSI-approved.

Why it matters

Open weights at the 428B scale are rare — most labs this size stay API-only. Releasing the shards with domestic Chinese silicon support front and center is a bid to own the self-hosted agent stack, not just the API bill. And the stealth run was free QA at planetary scale: OpenRouter’s telemetry let MiniMax ship launch-day fixes (time-to-first-token from 1.95s to 1.22s, throughput from 70 to 100 tokens per second, error rate halved from 1.4% to 0.7%) before the name was ever attached. Expect other labs to copy the playbook — anonymous chart-topping, then a name, then the weights.