Flash in the name, 428 billion parameters in the card. The name misleads. According to MiniMax, MiniMax-M3.1-Flash is a mixture-of-experts (MoE) model in which 23 billion parameters work on each token. It takes text, images and video, answers in text and holds a context of up to 1 million tokens. The purpose MiniMax lists: coding, agentic workflows, computer use and cybersecurity.
Fact
The facts: the MiniMax-M3.1-Flash page on Hugging Face is dated 25 October 2026. According to MiniMax, the model is a native multimodal mixture-of-experts model with about 428 billion parameters in total and 23 billion activated per token; inputs are text, image and video, output is text; the context window is up to 1 million tokens. Compared with MiniMax-M3 it has three changes: sparse attention throughout the model (the first three full-attention layers are replaced), quantization-aware training for the routed experts and the attention paths, and DSpark speculative decoding, which replaces the earlier EAGLE-like MTP approach; the release also includes the draft head. MiniMax's figures: 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. The license is listed as minimax-community.
They write about a generational leap in agentic tasks, coding and cybersecurity. The table on the same card is more modest. The columns are picked by MiniMax: DeepSeek V4.1 Flash-1017, DeepSeek V4.1 Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol and Opus 5. The model leads in two rows: Automation-Bench v1.0.6 with 62.8 and LOCA-Bench 256K with 60.3. On BankerToolBench it has 76.2, the same as Opus 5. It trails on DeepSWE v1.1 (73.7 against 77.0 for DeepSeek V4.1 Flash-1017), on Terminal Bench 4.0 (38.1 against 51.8 for Opus 5), on HLE without tools (43.4 against 55.6) and on OSWorld 2.0 (61.7 against 70.6).
The name says fast, the card says 428 billion.
Efficiency is the other half of the announcement, but without a number. The page claims that sparse attention, quantization-aware training and DSpark improve it, and does not say by how much. It gives no price for the MiniMax API. Usable context and throughput, it says, depend on GPU memory and the serving configuration, so a million tokens is a ceiling, not a guarantee.
The recommended frameworks are SGLang, vLLM and TokenSpeed; Transformers, KTransformers, unsloth and ATOM are still in progress. The scores are for the official weights, and third-party quantizations may differ. We do not go through the terms of the minimax-community license here: read them before you put the model in a product.
Memory first: serve it through vLLM or SGLang with the context you actually need, and see how far it holds.
Official primary sources
MiniMax - MiniMax-M3.1-Flash, Hugging Face, 25.10.2026The visual is generated code-art, no third-party images.