MiniMax-M3 at a glance.
From MiniMax's own pages, read Oct 20, 2026, so company-reported. A token is a word or part of a word. Prices can change; check the maker's page before you build on them.
- Made by
- MiniMaxBased in China
- Released
- Not stated
- What it is for
- “MiniMax-M3 is a native multimodal model with 1M context.”Hugging Face model card MiniMaxAI/MiniMax-M3, Oct 20, 2026 · in MiniMax's words
- Reads at once
- 1M tokensIts context window: how much text it can take in one go
- Takes in
- text
- Download
- Yes, open weightsLicense: minimax-community
- Size
- ~428B parameters, ~23B activeHugging Face model card MiniMaxAI/MiniMax-M3, Oct 20, 2026
$0.30per million tokens in$1.20per million tokens out
Prompts <=512k input tokens. Prompts >512k: $0.60 in / $2.40 out. Cache read $0.06 (<=512k) / $0.12 (>512k). Priority tier is 1.5x. Smaller MiniMax-M3.1-Flash: $0.15 in / $0.60 out, cache read $0.015.
MiniMax API platform: Pricing, Oct 20, 2026 · company-reported
Where it ranks.
Its position on each of the 21 tests the models page uses, by testers other than the maker. Epoch AI's and Vectara's scores are shown under their open licenses; for LMArena, Scale and Vals we give the position and link to their own leaderboard. Each list is a snapshot with its date.
| Test | Rank | Score |
|---|---|---|
| Epoch Capabilities IndexEpoch AI, Oct 20, 2026 | Not in the top 30 | |
| Vals IndexVals AI, Oct 19, 2026 | Not in the top 20 |
| Test | Rank | Score |
|---|---|---|
| LMArena WebDevLMArena, Oct 19, 2026 | Not in the top 20 | |
| SWE-Bench ProScale AI, Sep 22, 2026 | Not in the top 16 | |
| Vibe Code BenchVals AI, Oct 18, 2026 | Not in the top 20 | |
| LMArena coding chatsLMArena, Oct 15, 2026 | Not in the top 20 |
| Test | Rank | Score |
|---|---|---|
| LMArena textLMArena, Oct 15, 2026 | Not in the top 20 | |
| LMArena creative writingLMArena, Oct 15, 2026 | Not in the top 20 | |
| MultiChallengeScale AI, Feb 3, 2026 | Not in the top 15 |
| Test | Rank | Score |
|---|---|---|
| LMArena Agent ArenaLMArena, Oct 19, 2026 | Not in the top 20 | |
| Remote Labor IndexScale AI | Not in the top 16 | |
| Terminal-Bench 4.0Vals AI, Oct 19, 2026 | Not in the top 20 | |
| CUA-bench (games)Vals AI, Oct 13, 2026 | Not in the top 9 |
| Test | Rank | Score |
|---|---|---|
| LMArena mathLMArena, Oct 15, 2026 | Not in the top 20 | |
| Humanity's Last ExamScale AI | Not in the top 17 | |
| ProofBenchVals AI, Oct 12, 2026 | Not in the top 20 |
| Test | Rank | Score |
|---|---|---|
| LMArena visionLMArena, Oct 19, 2026 | Not in the top 20 | |
| LMArena documentsLMArena, Sep 13, 2026 | Not in the top 20 |
| Test | Rank | Score |
|---|---|---|
| GPQA DiamondVals AI, Sep 1, 2026 | #15 | On Vals AI |
| MMMU ProVals AI, Sep 1, 2026 | Not in the top 20 |
| Test | Rank | Score |
|---|---|---|
| Vectara hallucination testVectara, Oct 16, 2026 | Not in our copy of the list |
“Not in the top” means it isn't among the models in our copy of that list, which runs to the position shown; it may not have been tested. Vectara's test ranks by how rarely a model adds facts, so lower is better. Positions can be shared when scores tie. Epoch's range is its margin of uncertainty. Epoch AI and Vectara data: CC BY 4.0 and Apache-2.0.
Our note: Release date not stated on pages read; the Hugging Face repo was created 2026-06-02. The card calls it 'a native multimodal model' and says training spans 'text, image, and video', but we did not find an explicit list of accepted API input types, so only text is recorded.