M3 (batch)
MiniMax · released May 31, 2026
MiniMax-M3 is a multimodal foundation model from MiniMax. It supports text, image, and video inputs with text output, a 1M-token context window, and is suited for long-horizon agentic work, coding,...
Specification
- Context window
- 524K
- Max output
- —
- Knowledge cutoff
- Not stated
- Parameters
- 427B
- Licence
- other
- Serving providers
- 1
- Moderated
- No
Intelligence
45.4
93th percentile
Coding
58.6
Coding Index
Agentic
36.1
Agentic Index
Output speed
80 t/s
Median across providers
Latency
1.24s
Time to first token
Cost per task
$0.06
Estimated
Benchmarks
Where the score comes from
The Intelligence Index is a composite. These are the underlying evaluations this model was actually measured on.
Evaluation scores
Percentage correct · higher is better
- GPQA Diamond92.9%
- τ²-bench (Telecom)88.9%
- IFBench82.9%
- AA-LCR (long context)80.3%
- SciCode45.4%
- Terminal-Bench Hard42.4%
- Humanity's Last Exam39.0%
An evaluation missing from this list was not run for this model — it is not a zero.
View as table
| Evaluation | Score |
|---|---|
| GPQA Diamond | 92.9% |
| τ²-bench (Telecom) | 88.9% |
| IFBench | 82.9% |
| AA-LCR (long context) | 80.3% |
| SciCode | 45.4% |
| Terminal-Bench Hard | 42.4% |
| Humanity's Last Exam | 39.0% |
Against its peers
Intelligence Index · this model highlighted, nearest peers in grey
- Gemini 3.1 Pro PreviewGoogle47.7
- Qwen3.7 MaxQwen46.7
- GPT-5.3-CodexOpenAI45.5
- M3 (batch)MiniMax45.4
- MiniMax M3MiniMax45.4
- V4 ProDeepSeek45.3
- Kimi K2.6Moonshot AI45.1
- GPT-5.2OpenAI43.3
Peers are the models sitting closest on the Intelligence Index — the set you would realistically choose between.
View as table
| Model | Intelligence |
|---|---|
| Gemini 3.1 Pro Preview | 47.7 |
| Qwen3.7 Max | 46.7 |
| GPT-5.3-Codex | 45.5 |
| M3 (batch) | 45.4 |
| MiniMax M3 | 45.4 |
| V4 Pro | 45.3 |
| Kimi K2.6 | 45.1 |
| GPT-5.2 | 43.3 |
Percentile among all indexed models
Pricing
What it costs to run
List prices per million tokens, plus what one representative task works out to.
List price
- Input / 1M tokens
- $0.15
- Output / 1M tokens
- $0.6
- Cached input / 1M
- $0.03
- Blended 3:1
- $0.262
One task, estimated
$0.06
- Input tokens
- 50,000
- Output tokens
- 80,000
- Profile
- Reasoning
Estimated from list pricing: 50K input tokens plus 80K output tokens for reasoning models (25K for non-reasoning).
Serving providers
Speed and latency figures are medians across these providers, so a widely-served model reports a blend rather than any single endpoint.
Arena
Head-to-head generation quality
Elo from pairwise judgements, broken out by the kind of thing the model was asked to build.
- Overall Elo
- 1,285
- Win rate
- 54.0%
- Strongest at
- UI components
- Tournaments
- 1,175
Elo by category
Dot position on a 1,200–1,300 scale · Elo has no meaningful zero
- 1,285
- 1,284
- 1,274
- 1,262
- 1,261
- 1,216
- 1,237
- 1,226
Agent categories (full-stack apps, mobile apps) are only scored for models tested in agent harnesses.
View as table
| Category | Elo | Win rate |
|---|---|---|
| UI components | 1285 | 54.0% |
| Websites | 1284 | 54.4% |
| 3D scenes | 1274 | 53.9% |
| Game development | 1262 | 50.6% |
| Data visualisation | 1261 | 52.9% |
| SVG | 1216 | 49.7% |
| Mobile apps | 1237 | 55.4% |
| Full-stack apps | 1226 | 53.0% |