A point release of MiniMax-M3 that keeps the parameter budget and spends the work on serving cost. The model card describes the same natively multimodal MoE (text, image and video in, text out) at about 428B total and 23B active, and config.json matches M3's shape: 60 layers, hidden size 6144, 64 query heads and 4 KV heads, 128 routed experts with top-4 routing plus one shared expert, and a 1,048,576-token position limit. The card lists three changes, all aimed at inference. First, sparse attention throughout: the first three full-attention layers M3 kept are replaced, so MiniMax Sparse Attention now covers all 60 layers. Second, quantization-aware training of the routed experts and the attention paths. The checkpoint ships as MXFP8 with NVFP4 experts rather than as a post-hoc quantization. Third, a DSpark draft head replaces M3's EAGLE-style MTP, so MiniMax is shipping DeepSeek's speculative-decoding design in its own model. MiniMax says the result is "nearly three times faster" online inference, 200+ tokens/s peak and 170+ median, and 70% lower cost on agentic tasks than its predecessor. The API price is $0.15 in / $0.60 out / $0.015 cached input per 1M tokens.

Independent measurement backs the capability claim more than the self-reported table does. Artificial Analysis scored it 43 on Intelligence Index v4.3.2, against 29 for M3 on the same version — a 14-point gain inside a minor version that put M3.1-Flash just behind DeepSeek-V4.1-Flash-1017 (45) and ahead of the original V4.1-Flash (39). On v5.0 (October 27) it reads 29 against 19 for M3 and 32 for V4.1-Flash-1017, so the order holds. The vendor table, which uses a mix of third-party harnesses, shows the same pattern: BankerToolBench 76.2, DeepSWE v1.1 73.7, Automation-Bench v1.0.6 62.8, OSWorld 2.0 61.7, CyberGym 82.2, Agents' Last Exam 29.6, HLE (text-only, no tools) 43.4, and Terminal-Bench 4.0 38.1, where Opus 5 scores 51.8. The blog says the anonymous "Space Bunny" model, which served more than 200T tokens before launch, was an early version, and that an M3.1 Pro "will also arrive later".

The launch post said "model weights will be released later", but the HuggingFace repo was created on October 17 and holds 49 public, ungated safetensors shards under the MiniMax Community License (as for M3). AA has since relabelled it open weights and scores it 27.8 on its Openness Index. Nothing on the card says whether this is a continued-training checkpoint or a retrain. "Building on MiniMax-M3" and the identical shape suggest the former, so the entry records M3 as its base. Undisclosed: training tokens, the post-training recipe beyond QAT and sparsification, and any technical report (the card links only M3's arXiv paper).

Model Details

Architecture MOE
Active params 23B
Experts 128 (top-4)
Context window 1,048,576
AA Intelligence 29 was 43 on v4.3.2
License MiniMax Community License
Base model minimax-m3

Benchmark Scores

Benchmark Score Mode
BankerToolBench 76.2 —
DeepSWE v1.1 73.7 —
Automation-Bench v1.0.6 62.8 —
OSWorld 2.0 61.7 —
CyberGym 82.2 —
Agents' Last Exam 29.6 —
Terminal-Bench 4.0 38.1 —
HLE (text-only, no tools) 43.4 —
LOCA-Bench 256K 60.3 —
PostTrain Bench v1.1 39.3 —
NL2Repo-Bench 61.2 —
frontieropen-weightmoemultimodalagenticcodingefficiencyquantization

Related