Trending
Hot topicsCalmed down

MiniMax open-sources M3.1 Flash weights, with 428B total parameters

41 Articles16 Reporting sourcesUpdated 15 days ago

Learn about this first

On October 23, MiniMax open-sourced the model weights for MiniMax-M3.1 Flash, which has 428B total parameters and 23B active parameters per token. The weights are available for download on Hugging Face and ModelScope. The model previously operated under the anonymous codename Space Bunny on platforms including OpenRouter and OpenCode. MiniMax says real users processed more than 100 trillion tokens within 72 hours of its launch.

The model officially launched on MiniMax API, MPlan, and MiniMax Code on October 19. Official pricing per million tokens is RMB 1 for input, RMB 4 for output, and RMB 0.1 for cache reads; US dollar pricing is $0.15 for input, $0.60 for output, and $0.015 for cache reads, with no long-context surcharge. MiniMax says it is approximately 3 times as fast as M3 and reduces agent task costs by 70%. Artificial Analysis measured an Intelligence Index score of 42, 13 points higher than M3, with a cost per task of approximately $0.22.

After its open-source release, platforms including vLLM, SGLang, NVIDIA NeMo, Together AI, Fireworks, SiliconFlow, and Tenstorrent Cloud said they would provide inference, fine-tuning, or hosting support on day one. Unsloth released a 2-bit GGUF quantized version that is 45% smaller and can run locally with about 150GB of memory; its evaluation reports retaining 91.5% top-1 accuracy.

AI-generated from reports · updated 15 days ago

Latest developments

Overview of Local Quantization Options for the MiniMax M3.1 Flash 428B Model

Event progress27

  1. MiniMax M3.1 Flash Supports Running on More Local Hardware

    MiniMax: Overview of Local Quantization Options for the MiniMax M3.1 Flash 428B Model
  2. Unsloth releases support for running MiniMax-M3.1-Flash locally

    Unsloth: Unsloth releases quantized GGUF versions of MiniMax-M3.1-Flash; the 2-bit version can run locally with 150GB of memory
  3. MiniMax thanks 0G for supporting M3.1 Flash

    MiniMax: MiniMax thanks the 0G community for supporting M3.1 Flash
  4. MiniMax M3.1 Flash Releases Open Weights and Launches on Tenstorrent Cloud

    MiniMax: MiniMax M3.1 Flash launches on Tenstorrent Cloud
  5. Fireworks Launches a Serverless Preview of MiniMax M3.1 Flash

    MiniMax: MiniMax M3.1 Flash launches in Fireworks serverless preview
  6. MiniMax-M3.1-Flash Releases Open Weights

    SiliconFlow: MiniMax-M3.1-Flash launches on SiliconFlow on the first day of open-weight release
  7. MiniMax M3.1 Flash Receives Day-0 Support from TokenSpeed

    MiniMax: MiniMax thanks TokenSpeed for its day-one support
  8. MiniMax M3.1 Flash supports SGLang inference and RL training

    MiniMax: MiniMax M3.1 Flash receives support from SGLang and Miles
  9. MiniMax-M3.1-Flash Receives Day-0 Fine-Tuning Support in NVIDIA NeMo

    MiniMax: NVIDIA NeMo provides day-0 fine-tuning support for MiniMax-M3.1-Flash
  10. MiniMax M3.1 Flash open-sourced

    MiniMax: MiniMax opens the M3.1 Flash weights: 428B total parameters, with 23B activated per token
  11. vLLM Releases Day-0 Support for MiniMax-M3.1-Flash

    vLLM Official Blog: vLLM implements day-0 support for MiniMax-M3.1-Flash
  12. MiniMax announces Inferact as an API launch partner for M3.1 Flash

    MiniMax: MiniMax M3.1 Flash provides an API in partnership with Inferact
  13. MiniMax M3.1 Flash Integrates with SGLang Omni

    MiniMax: MiniMax M3.1 Flash teams up with SGLang Omni to build interruptible voice agents
  14. MiniMax M3.1 Flash model launches on Runware

    MiniMax: MiniMax M3.1 Flash launches on Runware
  15. MiniMax M3.1 Flash Tops the Search Efficiency Rankings

    MiniMax: MiniMax M3.1 Flash tops the search efficiency leaderboard
  16. Cinematic-quality creations made with MiniMax M3.1 Flash

    MiniMax: MiniMax M3.1 Flash cinematic creation showcase
  17. OpenClaw Supports MiniMax M3.1 Flash

    OpenClaw: MiniMax M3.1 Flash launches on OpenClaw
  18. Sharing the prompt for generating a river with MiniMax M3.1 Flash

    Tencent WorkBuddy: MiniMax M3.1 Flash generates a nonexistent river from a single prompt
  19. WorkBuddy Launches the MiniMax M3.1-Flash Model

    Tencent WorkBuddy: MiniMax M3.1-Flash launches on WorkBuddy
  20. MiniMax M3.1 Flash free for a limited time in Command Code

    MiniMax: MiniMax M3.1 Flash available free for a limited time
  21. MiniMax M3.1 Flash Launches on GMI Cloud

    MiniMax: MiniMax M3.1 Flash launches on GMI Cloud, focused on coding and agent workflows
  22. MiniMax M3.1 Flash available for free in Cline

    MiniMax: MiniMax introduces M3.1 Flash for free use in Cline, targeting long-horizon agentic coding scenarios
  23. MiniMax M3.1 Flash Launches on Ollama

    MiniMax: MiniMax M3.1 Flash launches on Ollama
  24. Artificial Analysis publishes MiniMax-M3.1-Flash evaluation results

    Artificial Analysis: Artificial Analysis evaluates MiniMax-M3.1-Flash: an Intelligence Score of 42, at less than half the cost of MiniMax-M3
  25. Arena Announces MiniMax-M3.1-Flash's Ranking in Code Arena

    Arena: MiniMax-M3.1-Flash joins the Code Arena WebDev leaderboard, ranking 12th among open-source models
  26. MiniMax-M3.1-Flash coming soon

    MiniMax: MiniMax-M3.1-Flash coming soon
  27. MiniMax M3.1 Flash Model Update Released

    MiniMax: Blog: MiniMax releases M3.1 Flash, nearly tripling inference speed and reducing usage costs by 70%

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.