Reportedly: MiniMax open-sources M3.1 Flash, a 428B-parameter multimodal model with 1M-token context
Reported by Zhidongxi · not yet confirmed by the company or a second independent outlet. We update this page when it is.
MiniMax released M3.1 Flash model weights on Hugging Face, a 428B total parameter model with ~23B active parameters per token. It supports text, image, and video input with up to 1M token context, and processed over 100 trillion tokens in 72 hours since launch, Zhidongxi reported.
China context
- Original name
- 稀宇科技
- Outside China
- Open weights · zhidx.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can download weights from Hugging Face and use vLLM or SGLang for inference; API access is available via Together AI.
- For investors
- The 70% cost reduction claim and aggressive pricing may pressure competitors' margins in the agent and coding API market.
Translated from Chinese. Quotes and facts link to the original sources.
MiniMax open-sourced M3.1 Flash, a multimodal model with 428B total parameters and ~23B active parameters per token. It supports text, image, and video input with up to 1M token context, targeting coding, agent, and computer-use tasks. The model processed over 100 trillion tokens for real users within 72 hours of its October 19 launch. Key techniques include full-model sparse attention, quantization-aware training (QAT), and power-aware scheduling, delivering nearly 3x faster inference than M3, 70% lower cost for agent tasks, and output speeds over 200 tokens per second. Ecosystem support includes Day 0 adaptation by Kunlunxin, NVIDIA NeMo fine-tuning, vLLM and SGLang inference, and Together AI availability. MiniMax is working with Huawei Ascend, T-Head, AMD, and Fireworks for first-day access. API pricing is 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.1 yuan per million cached read tokens. The model previously appeared as the anonymous 'Space Bunny' on OpenRouter, topping usage for 10 consecutive days with 56.7 trillion tokens processed.
M3.1 Flash uses full-model sparse attention, QAT, and power-aware scheduling to achieve 3x faster inference and 70% lower agent-task cost versus M3. It scores 73.7% on DeepSWE v1.1 (vs Claude Opus 5's 74.0%) and 62.8% on Automation-Bench (vs DeepSeek-V4.1-Flash-1017's 60.5%).
Developers outside China can now access a 428B-parameter open-weights model with 1M-token context at API prices roughly 1/33 (input) and 1/42 (output) of Claude Opus 5, directly lowering the cost of building coding and agent applications.
The model's low API pricing and open weights could reduce inference costs for enterprises building agentic and coding applications, but actual savings depend on workload and infrastructure.
Watch for confirmed Day 0 support from Huawei Ascend, T-Head, AMD, and Fireworks, and whether the model's performance holds on independent benchmarks beyond MiniMax's reported results.