MiniMax Opens M3.1-Flash Weights, a 428B Agent Model Beating Opus 5
MiniMax dropped the open weights for its 428B parameter multimodal MoE model, with 23B active per token, 1M context, and strong agentic benchmarks.
- MiniMax released open weights for M3.1-Flash, a 428B parameter multimodal MoE with 23B active per token.
- Repository is 252 GB across 48 safetensors shards plus a 2.3 GB DSpark speculative-decoding drafter.
- Supports text, image, and video input with a 1M token context window and 524K max output.
- Hosted API pricing is $0.15 per million input tokens and $0.60 per million output tokens.
- Leads Automation-Bench v1.0.6 at 62.8 and scores 82.2 on CyberGym; weaker on Terminal Bench and HLE.
- Community license permits commercial use until $20M annual revenue, then requires MiniMax approval.
MiniMax has released the weights for M3.1-Flash on Hugging Face, four days after launching its hosted API. The native multimodal mixture-of-experts model targets coding agents, long-running computer-use tasks, and cybersecurity workflows.
The release combines a 1 million-token input window with text, image, and video support. MiniMax reports strong results on several agent benchmarks, giving developers another option for running large agent models on their own infrastructure.
Inside the 428B checkpoint
| Specification | M3.1-Flash |
|---|---|
| Total parameters | 428 billion |
| Active parameters per token | 23 billion |
| Repository size | 252.4 GB |
| Weight files | 48 shards |
| Speculative-decoding model | 2.3 GB |
| Inputs | Text, images, and video |
| Output | Text |
| Maximum input | 1,000,000 tokens |
| Maximum output | 524,288 tokens |
A mixture-of-experts router activates 23 billion parameters for each token while retaining 428 billion parameters of total capacity. That design reduces computation per token, though servers must still store and distribute the full checkpoint.
Three changes target inference cost
M3.1-Flash retains the mixture-of-experts architecture used by MiniMax M3 and changes its attention, quantization, and speculative-decoding systems.
- Sparse attention in every layer: MiniMax replaced the first three full-attention layers with MiniMax Sparse Attention. Sparse attention processes a selected subset of earlier tokens, reducing the work required for very long sequences.
- Quantization-aware training: the routed experts and attention paths were quantized during training. The checkpoint consequently includes
mxfp8, an 8-bit microscaling format, and other 8-bit tensors. - DSpark speculative decoding: a dedicated draft model replaces the EAGLE-style multi-token prediction system used in M3. MiniMax includes the drafter with the checkpoint.
During speculative decoding, the smaller model proposes several tokens and the main model verifies them together. Accepted proposals increase generation throughput, with the actual gain depending on acceptance rates, prompts, hardware, and serving-engine support.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
