MiniMax M3 — the weights are out, the license is the catch

MiniMax's flagship native-multimodal MoE, M3, shipped its open weights to Hugging Face on ~June 7–11 (after the June 1 API launch) — resolving the item we deferred on June 5. It is ~428B total / ~23B active, 1M context, text + image + video in. Two things keep it off the local picks: it is big-iron-class at 428B, and it ships under the custom "MiniMax Community License," not an OSI-permissive one. So it lands as a fast-take and a cloud-side comparator, not a planner recommendation.

Verdict: Open weights finally landed — but a non-OSI "community" license and 428B total keep it a frontier comparator, not a local pick

The take

The facts, verified against the Hugging Face model card and the LICENSE file: MiniMax-M3 is ~428B total parameters / ~23B activated, native multimodal (text + image + video in), 1M context, uploaded to `MiniMaxAI/MiniMax-M3` on 2026-06-07. The license in the repo is the "MINIMAX COMMUNITY LICENSE" — a custom license, not Apache or MIT; it permits free use but adds conditions (commercial-use and attribution terms in the family of MiniMax's earlier contested licensing). That is the same posture as our existing MiniMax M2 entry, which we already flag as commercial-license-gated rather than cleanly open.

Why it matters: M3 landing as downloadable weights is genuinely notable — a frontier-class native-multimodal MoE with a 1M context window, openly downloadable, is rare. But "open weights" and "open license" are not the same thing, and the community reaction has centered on exactly that: the license is restrictive and ambiguous enough that several local-AI practitioners called it a non-starter for real use. We hold OSI-permissive (Apache / MIT) as the bar for a planner pick precisely to avoid recommending something a reader can't actually deploy freely.

Our call: no planner-pick change, and no new model detail page this sweep. At 428B total, M3 is datacenter / heavy-multi-GPU class regardless of license — it does not fit the planner's hardware tiers as a usable local model. We track it the way we track DeepSeek V4-Pro, Kimi K2.6, and Command A+: a frontier comparator that shows where the open(-ish) ceiling is. If MiniMax relicenses M3 under clean OSI terms or ships a much smaller variant, we will revisit. For now: the weights being out is the news; nothing you run locally changes.

Where this fits

Models: MiniMax M2.5 / M2.7 · Kimi K2.6 · Command A+ (218B-A25B) · DeepSeek V4-Pro

Hardware: NVIDIA DGX Spark · Dual RTX 5090 · Mac Studio M3 Ultra 96 GB

Sources

Next step

Try this in the planner→