M3: Frontier Coding, Not Frontier Freedom

Setup

I loaded MiniMax M3 into my harness and pointed it at a real repository — not a benchmark, a codebase. 428 billion parameters total, 23 billion active per token. A mixture-of-experts model with a 1-million-token context window and a 512K guaranteed floor. Most Western developers have never pulled these weights. That is a geography problem, not a quality problem.

The specs are not marketing. M3 runs on a new attention design MiniMax calls MSA — MiniMax Sparse Attention — documented in a dedicated arXiv paper (2606.13392) co-authored with NVIDIA and three Chinese universities. Blockwise sparse attention layered on grouped query attention, with an index branch that scores KV blocks and keeps only the top-k per group. That is the whole trick for staying fast when the context runs long.

Observation

At 1M tokens, MiniMax reports per-token compute cut to one-twentieth of M2 — more than 9× faster prefill, more than 15× faster decode. The paper’s raw H800 kernel numbers are more conservative and more honest: 14.2× prefill, 7.6× decode. Different measurements, both first-party, both pointing the same direction.

On its own model card, M3 scores 59.0% on SWE-Bench Pro and 80.5% on SWE-Bench Verified. MiniMax also claims it beats GPT-5.5 and Gemini 3.1 Pro on the coding suites. That comparison is MiniMax’s, not an independent leaderboard’s, so I read it as a claim, not a fact.

Throughput is where the numbers go soft. MiniMax publishes no official tokens-per-second figure. Artificial Analysis measured about 85.7 t/s against MiniMax’s own hosted API; third-party providers land anywhere from 37 to 101 depending on who is serving it. Fast enough. Unproven at a single number.

Pricing is public and specific. $0.30 per million input tokens, $1.20 per million output — at or below 512K context. Cross 512K and it doubles to $0.60 and $2.40. Cheap for frontier-class coding.

Then the catch. M2 shipped under plain MIT. M3 does not. It ships under something called the MiniMax Community License. Commercial use requires you to display “Built with MiniMax M3” on your product. Earn more than $20M a year in relevant revenue and you need written authorization before you ship. That is not open the way M2 was open. It is open-weight with a leash.

Take

Here is the position, sharpened past “it’s good and the West ignores it.” M3 is the correct model for exactly one job: long-context agentic coding above 100K tokens. When your agent has to hold an entire repository — cross-file refactors, a dependency graph, a week of session history — in a single window and still move, the MSA architecture and the 512K-guaranteed floor earn their keep, and the price makes it viable to run all day. That is a named use case, not a vibe.

But run it knowing what you signed. The weights download. The license is not MIT. And the moment your revenue crosses the threshold, the model under your product needs a permission slip from a lab in Shanghai. Load M3 for the long-context work it wins. Read the license before you build a business on it. Open weights and open terms are two different freedoms — M3 hands you the first and quietly takes back part of the second.