MiniMax-M3
MiniMax's open-weight model from June 2026 with a 1M-token context window, text and image input, and low per-token prices. It now trails the newer open models on most tests but remains one of the cheapest options here.
MiniMaxLanguage modelReleased June 2026Data checked October 18, 2026
This model is not in the current people's rankings: it is either new and still collecting votes, or ranked below the top 12. See the AI leaderboard for the models people rate highest.
On paper
Fixed-test rankings from the benchmarks tab: Speed #10 (tied), Reasoning #20.
Strengths and weaknesses
- Low price ($0.30 / $1.20 per million tokens for prompts up to 512K tokens)
- Strong instruction following (82.9% on IFBench as measured by Artificial Analysis)
- Open weights on Hugging Face
- Low Artificial Analysis Intelligence Index (29.2) and a near-zero Terminal-Bench 4.0 score (2.0% as run by Artificial Analysis)
- Uses a custom MiniMax community license rather than a standard open-source license
Best for
- Cheap long-context tasks
- Following detailed instructions
Every score we have
Each bar compares this model with the best result any model on this page has on the same test.
- LMArena text arena (human votes)1,441.3Arena score 1441.3 (95% CI 1437.5 to 1445.2); 63,973 votes; entry "minimax-m3"; position 96 of 416 on the full board
- LMArena creative writing1,405.5Arena score 1405.5 (95% CI 1398.8 to 1412.1); 11,713 votes; entry "minimax-m3"; position 101 of 414 on the full board
- LMArena WebDev (coding)1,482.5Arena score 1482.5 (95% CI 1476.9 to 1488.0); 19,499 votes; entry "MiniMax-M3"; position 64 of 144 on the full board
- LMArena vision1,235.5Arena score 1235.5 (95% CI 1229.1 to 1242.0); 19,026 votes; entry "minimax-m3"; position 69 of 165 on the full board
- LMArena Agent Arena-0.083IPS score -0.0831 (95% CI -0.0941 to -0.0721); 42,848 sessions; entry "MiniMax-M3"; position 47 of 52 on the full board
- Artificial Analysis Intelligence Index29.229.2 on the Artificial Analysis Intelligence Index v4.3.2 (AA shows 29), setting "MiniMax-M3"
- Humanity's Last Exam (Artificial Analysis)39.0%39.0% on Humanity's Last Exam (AA, 2,158 text-only questions), setting "MiniMax-M3"
- GPQA Diamond, graduate-level science92.9%92.9% on GPQA Diamond (AA; no longer part of the AA index since v4.2), setting "MiniMax-M3"
- Terminal-Bench 4.0 (Artificial Analysis run)2.0%2.0% on Terminal-Bench 4.0 as run by AA (mini-swe-agent harness), setting "MiniMax-M3"
- Long-context reasoning (AA-LCR)83.0%83.0% on AA-LCR (long-context reasoning), setting "MiniMax-M3"
- MMMU-Pro, image understanding78.6%78.6% on MMMU-Pro (image understanding, AA), setting "MiniMax-M3"
- IFBench, following instructions82.9%82.9% on IFBench (instruction following, AA), setting "MiniMax-M3"
- AA-Omniscience, accuracy minus hallucination1.41.4 on the AA-Omniscience Index (knowledge accuracy minus hallucination, range -100 to 100), setting "MiniMax-M3"
- Output speed, tokens per second93 t/s93 output tokens per second (AA median), row "MiniMax-M3"
- Cost per Artificial Analysis test task$0.51$0.51 to run one AA Intelligence Index task, row "MiniMax-M3"
Compare with
- Claude Opus 5.5Anthropic's flagship for long-running agentic coding and knowledge work, and the model Anthropic suggests starting with for most jobs.
- Claude Sonnet 5.5Anthropic's mid-priced model, which Anthropic describes as its best mix of speed and intelligence.
- Claude Fable 5.1Anthropic's top tier for the hardest reasoning and long-horizon agent work, which Anthropic recommends when Opus 5.5 still falls short.
- Claude Haiku 5.5Anthropic's smallest and fastest current model, built for high-volume, latency-sensitive jobs like classification, extraction and routing.