MiniMax-M3.1-Flash

What it is

MiniMax-M3.1-Flash is a large language model from MiniMax built on a sparse mixture-of-experts (MoE) design. In an MoE model the network is split into many separate expert subnetworks, and each token is routed through only a small subset of them, so the number of parameters that must be stored is much larger than the amount of computation spent per token. For M3.1-Flash the model card gives about 428B total parameters with 23B activated.^[1] The model accepts text, image and video input and supports contexts of up to 1M tokens.^[1] Its weights were posted to Hugging Face on 17 October 2026,^[2] under a minimax-community license, and reasoning effort can be set anywhere from low to max.^[1]

What changed from M3

M3.1-Flash reads as a revision of the earlier M3 line rather than a new architecture. The card describes three changes.^[1]

Attention. M3 kept full attention in the first three layers of the stack; M3.1-Flash removes that exception, so MiniMax Sparse Attention is used throughout the model.^[1] Sparse attention means each token is compared against a selected subset of positions rather than against every position. Since the number of comparisons sets the cost of attention, cutting comparisons is what makes very long contexts affordable.

Memory. Quantization-aware training covers the routed experts and the attention paths, in order to reduce memory use.^[1] Quantization-aware training means the model is trained using the reduced numerical precision it will later run at, so the low-precision weights it ends up with are closer to what training actually optimized for. Applying it to the routed experts matters most for memory, because those experts hold most of the 428B parameters.

Decoding. For generation, DSpark speculative decoding replaces the EAGLE-like multi-token prediction used in M3.^[1] Speculative decoding drafts several tokens cheaply and then has the full model check them, so a larger share of the work is done in parallel. The change concerns how the model is served rather than what it was trained to do.

Reported scores

The card gives DeepSWE v1.1 at 73.7%, OSWorld 2.0 at 61.7, CyberGym at 82.2 and BankerToolBench at 76.2.^[1] It also lists Terminal Bench 4.0 at 38.1, LOCA-Bench 256K at 60.3 and NL2Repo-Bench at 61.2.^[1] On the knowledge-and-reasoning side it reports HLE without tools at 43.4 and CharXiv RQ at 82.5,^[1] and it records GDP.pdf AA at 20.8 and Agents' Last Exam at 29.6.^[1]

The spread is wide because the tests measure different things: some examine work over long context, some examine whether a task can be carried through end to end in a terminal or a desktop environment, and some examine whether a research-level question can be answered without tool help. Because the tests differ that much, the numbers are best read one at a time rather than folded into a single ranking.

A discrepancy on the model card

The safetensors metadata on the card lists 220B parameters, which does not match the roughly 428B total given in the card text.^[1] Until that is clarified, the two figures should not be used interchangeably.

Where it sits

Taken as a whole, the release is an iteration: the same sparse MoE family, the same multimodal input, but with the full-attention exception removed, quantization-aware training extended to the experts and attention paths, and a new speculative decoding scheme. The most concrete combination for users is a 1M-token context^[1] together with attention that is sparse throughout,^[1] so that long inputs no longer carry the cost of full attention.