In short
How they differ.
Worked out from the figures below. Prices and specs are the makers' own, read Oct 20, 2026.
- MiniMax-M3 reads more at once: a context window of 1M tokens against 262K.
- Both can be downloaded and run on your own machine.
- No independent board we follow ranks both models yet.
Side by side
The specs.
| MiniMax-M3 | Qwen3.8-27B | |
|---|---|---|
| Made by | MiniMax · China | Alibaba (Qwen) · China |
| Released | Not stated | Aug 5, 2026 |
| Price per 1M tokens | $0.30 in / $1.20 out | Free to download; hosts set prices |
| Context window | 1M | 262K |
| Takes in | text | text, image, video |
| Download and run it yourself | Yes (minimax-community (license tag shown on Hugging Face)) | Yes (Apache 2.0) |
| Size | ~428B parameters, ~23B active | 27B (dense) parameters |
| Maker says | “MiniMax-M3 is a native multimodal model with 1M context.” | “a native vision-language model that understands images and videos, with flexible thinking control” |
| Sources | Prices, Oct 20, 2026 · Description | Prices, Oct 20, 2026 · Description |
Independent tests
Where they place.
Position on each board we follow (1 is best). Epoch AI's and Vectara's scores are shown; the other boards are linked for their own figures.
| Board | MiniMax-M3 | Qwen3.8-27B |
|---|---|---|
| GPQA Diamond (run by Vals AI, archived)Vals AI, Sep 1, 2026 | #15 | Not listed |
More comparisons