MiniMax M3: 1M Context and Native Video at $0.30/M
MiniMax released MiniMax M3 on June 1: a 1M-token context model with native image and video input, near-frontier coding claims, and a price of $0.30 per million input tokens and $1.20 per million output. MiniMax says the weights will follow within about ten days.
Long context alone is no longer rare in open weights; DeepSeek's V4 family has offered 1M tokens under MIT since April. What M3 adds is native multimodality at that context length and that price. If the weights ship as promised, it is the strongest case yet for an open model that reads code, documents, screenshots, and video in one pass.
What M3 is
- Inputs: text, image, and video, trained multimodally from the start rather than with a vision adapter bolted on afterwards.
- Context: up to 1M tokens, with 512K guaranteed.
- Attention: MiniMax Sparse Attention (MSA). MiniMax reports more than 9x faster prefill and more than 15x faster decoding at 1M tokens, and about 1/20 the per-token compute of its previous generation.
- Pricing: $0.30/$1.20 per million tokens up to 512K of input; $0.60/$2.40 from 512K to 1M.
The benchmark claims
From MiniMax's launch post, all self-reported:
- SWE-bench Pro: 59.0%
- Terminal-Bench 2.1: 66.0%
- MCP Atlas: 74.2%
- BrowseComp: 83.5
- PostTrainBench: 0.37, behind Claude Opus 4.7 (0.42) and GPT-5.5 (0.39)
The long-horizon demos are more interesting than the tables. In one, M3 spent nearly 12 hours reproducing an ICLR 2025 Outstanding Paper, producing 18 commits and 23 figures. In another, it spent about 24 hours optimizing an FP8 matrix-multiply kernel: 1,959 tool calls, 147 benchmark submissions, and hardware utilization raised from 7.6% to 71.3%, a 9.4x speedup. These are demos chosen by the vendor, but they test sustained autonomous work, which short benchmarks don't.
Artificial Analysis puts M3 at 55 on its Intelligence Index, just ahead of Kimi K2.6 and MiMo-V2.5-Pro (both 54). That makes it the leading open-weight model on that index once the weights are public. One caveat stands out: on AA-Omniscience, M3 attempts only 30.9% of questions, the lowest of its peers. It hallucinates rarely because it often declines to answer. Whether that is good depends on whether your application prefers "I don't know" to a guess.
How it compares
| Model | LMRank | Input $/M | Output $/M | Image/video in | Weights |
|---|---|---|---|---|---|
| Claude Opus 4.7 | 9.6 | $5.00 | $25.00 | Image | Closed |
| GPT-5.5 | 9.4 | $5.00 | $30.00 | Image | Closed |
| DeepSeek V4 Pro | 9.0 | $0.435 | $0.87 | No | MIT |
| MiniMax M3 | 8.5 | $0.30 | $1.20 | Image and video | Promised |
Against DeepSeek V4 Pro, M3 is slightly cheaper on input, more expensive on output, lower on LMRank, and the only one of the two that reads images and video. Against the closed flagships, it is roughly 17x cheaper on input and 20x or more on output.
What to watch
- The license. MiniMax's previous model, M2.7, shipped under a commercially restricted license. "Open weights" can mean very different things; read the terms before you plan on self-hosting.
- Video in practice. Native video input at 1M tokens is the headline capability. Test it on your own footage; vendor demos pick clean inputs.
- Refusal rate. The low attempt rate on knowledge questions will show up as unhelpful answers in some products.
Update: the weights landed on Hugging Face on June 7 under a custom MiniMax community license, more permissive than M2.7's.
Takeaway
If your workload mixes long documents or codebases with screenshots or video, M3 is the first open-weight model worth testing for it, at a price that makes large-scale testing cheap. For text-only coding, DeepSeek V4 Pro remains the stronger open option.
Related: Best Open-Weight Models · Multimodal Models