MiniMax M3
MiniMaxMiniMax's M3 flagship, launched 1 June 2026 with MiniMax Sparse Attention, a one-million-token context and native multimodality. It was API-first at launch, with weights subsequently published on Hugging Face alongside vLLM and SGLang deployment guides. The training data and pipeline are not released.
Consolidated 2026-07-29 from minimax-m2-5; openness follows M3, the governing release. Capability is the one axis that does not describe M3, because MiniMax has published no benchmark table for it, so the band still rests on the M2.5 card. Verified 2026-08-13 via the M3 and M2.5 model cards.
Openness
3 high confidence- weights
- open(on HF
- data
- closed
- code
- partial(inference only
- license
- minimax-community(custom, see HF LICENSE)
Weights are distributed under the custom minimax-community license over a closed corpus, so 3. MiniMax-M3 is the current release and governs this axis. The capability score, by contrast, is carried from M2.5, because MiniMax has published no benchmark table for M3 - it is the one axis here not sourced to the current release.
- https://huggingface.co/api/models/MiniMaxAI/MiniMax-M3 recorded 2026-08-13
Public ungated weights repo (safetensors/MoE), `license: other` with `license_name: minimax-community`, `gated: false`, created 2026-06-02, 169,325 downloads.
- https://huggingface.co/MiniMaxAI/MiniMax-M3 recorded 2026-08-13
M3 model card with a bundled LICENSE and vLLM / SGLang deployment guides; no training corpus named and no training pipeline linked, which is what holds data at closed and code at partial.
- https://venturebeat.com/technology/minimax-m3-debuts-eclipsing-gpt-5-5-and-gemini-3-1-pro-on-key-benchmark-performance-for-just-5-10-of-the-cost recorded 2026-06-04
VentureBeat piece on the MiniMax M3 debut; the host answers HTTP 429, so it is not a live source for this axis.
Adoption
3 high confidence169,325 Hugging Face downloads in the trailing 30 days for the declared artifact MiniMaxAI/MiniMax-M3, the release the product records as currently governing, which bands at level 3 (100K-1M) on the model adoption scale. MiniMax is additionally served through OpenRouter and offered free inside coding tools such as Kilo Code, which a download count does not capture.
- https://huggingface.co/api/models/MiniMaxAI/MiniMax-M3 recorded 2026-08-13
169,325 downloads in the trailing 30 days for MiniMaxAI/MiniMax-M3, the single declared artifact
- https://huggingface.co/MiniMaxAI/MiniMax-M2.5 recorded 2026-06-04
~693,331 downloads last month on the official HF model page
Capability
5 medium confidenceFrontier-tier among open-weights agentic chat models - SWE-bench Verified 80.2%, with strong reasoning and browsing scores - and at very low serving cost. The figures are MiniMax-reported on the Hugging Face card and not independently audited, though they are consistent with peer open models such as Kimi K2.6 and DeepSeek-V4. The real limit is that all five come from the M2.5 card while M3 is the release that governs the family, and MiniMax has published no comparable table for M3.
- https://huggingface.co/MiniMaxAI/MiniMax-M2.5 recorded 2026-08-13
MiniMax-reported SWE-bench Verified 80.2%, SWE-bench Pro 55.4%, GPQA-D 85.2%, AIME25 86.3% and BrowseComp 76.3% on the official HF card
Verified 2026-08-13