MiniMax releases M3.1-Flash: 428B-parameter multimodal MoE with 1M-token context and weights on Hugging Face
MiniMax has published MiniMax-M3.1-Flash on Hugging Face: a native multimodal mixture-of-experts model with about 428B total parameters, 23B activated per token, and a context window of up to 1M tokens. It builds on MiniMax-M3 with sparse attention, quantization-aware training and DSpark speculative decoding.
MiniMax has released MiniMax-M3.1-Flash, a native multimodal mixture-of-experts (MoE) model with approximately 428B total parameters and 23B activated per token, with weights published on Hugging Face. It accepts text, image and video inputs, produces text output, and supports a context window of up to 1,000,000 tokens. MiniMax positions it for coding, agentic workflows, computer use and cybersecurity.
Architecture and changes from M3
M3.1-Flash keeps the core MoE architecture of MiniMax-M3 and makes three changes:
- Sparse attention throughout: MiniMax Sparse Attention (MSA) replaces the first three full-attention layers.
- Quantization-aware training: the routed MoE experts and the attention paths are quantized during training, which MiniMax says reduces memory footprint and improves serving efficiency.
- DSpark speculative decoding: DSpark replaces the earlier EAGLE-like MTP approach. The release includes the draft head for use with compatible serving implementations.
The Hugging Face page lists a safetensors model size of 220B params, with tensor types F32, BF16, F8_E4M3 and U8. The model card itself specifies approximately 428B total parameters.
Benchmarks (reported by MiniMax)
MiniMax claims 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0, 29.6 on Agents' Last Exam and 82.2 on CyberGym. Its comparison table includes DeepSeek V4.1 Flash-1017, DeepSeek V4.1 Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Selected results for M3.1-Flash, then GPT-5.6 Sol and Opus 5:
| Benchmark | M3.1-Flash | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|
| Automation-Bench v1.0.6 | 62.8 | 45.8 | 50.3 |
| BankerToolBench | 76.2 | 74.6 | 76.2 |
| Terminal Bench 4.0 | 38.1 | 39.9 | 51.8 |
| LOCA-Bench 256K | 60.3 | 60.0 | 53.7 |
| OSWorld 2.0 | 61.7 | 65.7 | 70.6 |
| DeepSWE v1.1 | 73.7 | 73.0 | 74.0 |
| NL2Repo-Bench | 61.2 | 56.8 | 75.3 |
| CyberGym | 82.2 | 83.6 | not reported |
| HLE (without tools, text-only subset of 2,158 questions) | 43.4 | 49.5 | 55.6 |
Other M3.1-Flash scores: GDP.pdf AA 20.8, PostTrain Bench v1.1 39.3, SWE Refactor Bench 23.0, BabyVision (with tools) 91.0, CharXiv RQ 82.5.
Harnesses differ by benchmark. BankerToolBench used OpenCode 1.17.18 with a MiniMax-M2.7 judge. Terminal Bench 4.0 used Codex 0.149.1. DeepSWE v1.1 used mini-SWE-agent 2.4.5 at 512K context. CyberGym used Claude Code 2.1.263 at 1M context. CharXiv RQ used a GPT-5.5 judge. Blank cells in MiniMax's table mean no official result, not zero. MiniMax notes that results are for the official weights and third-party quantizations may differ.
Deployment
The weights can be pulled with hf download MiniMaxAI/MiniMax-M3.1-Flash. MiniMax recommends SGLang, vLLM and TokenSpeed for serving, with Transformers, KTransformers, unsloth and ATOM integrations in progress. Multi-chip deployment is covered through FlagOS. The model is also available through MiniMax Agent and the MiniMax API. API pricing is not yet disclosed.
Recommended sampling settings are temperature=1.0 and top_p=0.95. Reasoning depth is set through reasoning_effort in the chat template, with options low, medium, high, xhigh and max. Usable context length and throughput depend on GPU memory and serving configuration.
As of 2026-10-23, the Hugging Face page showed 357 downloads in the last month, 3 finetunes and 14 quantizations in the model tree, and no Inference Provider deploying the model.
What this means
The 23B active-parameter count sets the per-token compute cost for self-hosting, while the full expert set determines memory needs. Quantization-aware training of the experts and attention paths is aimed at the memory side, though MiniMax gives no throughput or latency figures for the efficiency claims.
On MiniMax's own table, M3.1-Flash is strongest on agentic and tool-use tasks: it is highest on Automation-Bench and LOCA-Bench 256K among the listed models and ties Opus 5 on BankerToolBench at 76.2. It trails Opus 5 on Terminal Bench 4.0, OSWorld 2.0, NL2Repo-Bench and HLE. Because harnesses, context lengths and judges vary by benchmark, and the numbers come from MiniMax, independent replication on a common setup will matter for any head-to-head comparison.
Related Articles
Shanghai AI Lab releases Sage-A1.5, a 35B MoE agentic model with 256K native context
Shanghai AI Laboratory has released Sage-A1.5, a 35B-parameter Mixture-of-Experts agentic model with a native 256K context window, extendable to 512K via YaRN. The lab claims it matches or beats larger frontier models on several agentic and scientific benchmarks, including 95.1 on GAIA. Pricing is not yet disclosed.
Tencent releases WeVisOne-VL-2B, a 2B-parameter model for object detection, segmentation and alpha matting
Tencent's WeChat Vision team released WeVisOne-VL-2B, a 2B-parameter vision-language model built on Qwen3-VL. It handles object-level and pixel-level perception, including referring segmentation and alpha matting, through structured next-token prediction and a 9.5M-parameter pixel decoder.
inclusionAI releases Ling-3.0-flash-Sante, a 124B healthcare model with 5.1B active parameters
inclusionAI has released Ling-3.0-flash-Sante, a healthcare-specialized model built on Ling-3.0-flash with 124B total and 5.1B active parameters per token. Weights are on Hugging Face and ModelScope, and training extends to 256K tokens. The company labels it a research model, not a medical device.
Xiaomi open-sources Xiaomi-Lens-0, a 4B multimodal model that outputs pixel-level mask tokens mid-answer
Xiaomi's SeerRay-Lab has open-sourced Xiaomi-Lens-0, a 4B multimodal model built on Qwen3.5-4B. The model emits mask tokens for pixel-level image regions in the middle of an answer. The lab also released a benchmark called LensBench.
Comments
Loading...