HeyHomees Lab Notes · No. 9 · MiniMax · Published October 26, 2026
By HeyHomees Newsroom ·
In 30 seconds
- Native multimodal MoE model (text, image, video) with 1M token context window.
- Runs 23B activated parameters per token out of ~428B total.
- Scores 76.2 on BankerToolBench and 82.2 on CyberGym, showing strong agentic and cybersecurity performance
What is it?
MiniMax-M3.1-Flash is a native multimodal mixture-of-experts (MoE) model designed for coding, agentic workflows, computer use, and cybersecurity. It accepts text, image, and video inputs and produces text output. The model supports a context window of up to 1,000,000 tokens.
How does it work?
The model builds on MiniMax-M3 and adds three main changes: sparse attention throughout the model replaces the first three full-attention layers; quantization-aware training quantizes routed MoE experts and attention paths; and DSpark speculative decoding replaces the earlier EAGLE-like draft head, providing a draft head for compatible serving implementations.
How well does it do?
MiniMax-M3.1-Flash reports 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and 82.2 on CyberGym, indicating a generational leap in agentic, coding, and cybersecurity tasks.
What does it need to run?
The model has approximately 428B total parameters with 23B activated per token. Recommended serving frameworks include SGLang, vLLM, and TokenSpeed, and multi-chip deployment is supported via FlagOS. Recommended inference settings are temperature=1.0, top_p=0.95, and a reasoning_effort flag for deeper reasoning.
By the numbers
- Total parameters: Approximately 428B total
- Activated parameters per token: 23B activated per token
- Context window: Up to 1,000,000 tokens
The honest read
The reported scores (e.g., 76.2 on BankerToolBench, 82.2 on CyberGym) come from MiniMax’s own evaluation and are therefore self-reported; the source does not provide external verification or comparisons to other models. The benchmarks cover a range of tasks but many entries are blank for other models, so relative performance cannot be fully assessed. The model’s large parameter count and sparse activation reduce memory use, yet the source gives no concrete throughput or hardware utilization numbers, so efficiency claims are not quantified beyond the architectural description. The context window of up to 1M tokens is a specification rather than a measured capability; actual usable length will depend on GPU memory and serving configuration, as noted by the authors. Scores such as 73.7% on DeepSWE v1.1 are percentages, while others are raw points, making direct comparison across benchmarks difficult.
What we don't know yet
- Exact GPU memory requirements for full 1M token context.
- Throughput or latency numbers for inference.
- How the model performs on non-agentic tasks compared to other state-of-the-art models.
Sources
- Model card, Hugging Face (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.