MiniMax M3
Language Available Full comparison ↗ minimax/m3 · by MiniMax
· mixture-of-experts
Pricing — 1 offering(s)
About this offering: MiniMax M3 (API identifier: "MiniMax-M3"). Open-weight native-multimodal frontier model with 1M token context window. Pricing is tiered by context length: inputs ≤512k tokens use standard rates ($0.30/$1.20 per 1M); inputs >512k tokens are billed at double the rate ($0.60/$2.40) — currently limited availability, contact MiniMax sales. Prices shown include a permanent 50% promotional discount from list price. Priority-tier service (1.5× standard) available via service_tier:priority parameter. International API endpoint: api.minimax.io/v1/text/chatcompletion_v2.
Input tokens (≤512k context)
- $0.30 / 1M tokens (input) Current 2026-05-15 → present
Cached input tokens (≤512k context)
- $0.06 / 1M tokens (cached input) Current 2026-05-15 → present
Output tokens (≤512k context)
- $1.20 / 1M tokens (output) Current 2026-05-15 → present
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how MiniMax M3 fits into a cost-aware routing setup
See how →Capability profile
How MiniMax M3 rates across core capability dimensions, with the task-level evidence behind each rating.
MiniMax's own API docs name "agentic reasoning" as one of M3's primary capabilities (platform.minimax.io, already cited in this registry's capabilities.agentic_tool_use flag) — vendor-positioning based; no independent reasoning-specific benchmark (GPQA/MATH) was found to confirm this at time of writing.
SWE-bench Verified (500 human-verified GitHub issues; patch must pass the repo's own tests; no file-location hints): resolves 80.5% — fails roughly 1 issue in 5. Vendor-reported (MiniMax Hugging Face model card, Claude Code scaffold; not independently confirmed on a third-party leaderboard).
MiniMax's own API docs list `tools` and `tool_choice` parameters as "Fully supported" for MiniMax-M3 — same source already cited in the registry pricing entry's capabilities.agentic_tool_use flag.
1M-token context window, purpose-built via the MSA architecture (see architecture notes).
Vendor-published relative figure: 9x faster prefill / 15x faster decode versus MiniMax's own prior generation (M2) at 1M context (Hugging Face model card) — a within-family comparison, not benchmarked against other labs' models, so treat as directional rather than an absolute cross-model claim.
Coding benchmarks
View the full coding benchmark leaderboard →| Benchmark | Score | Harness | Source |
|---|---|---|---|
| swe-bench-verified ↗ | 80.5% 2026-06 | claude-code | 🏢 vendor ↗ |
| swe-bench-pro | 59.0% 2026-06 | claude-code | 🏢 vendor ↗ |
| terminal-bench-2-1 | 66.0% 2026-06 | terminus-2 | 🏢 vendor ↗ |
| terminal-bench-4-0 | 1.0% 2026-09 | mini-swe-agent | 📊 leaderboard ↗ |
Operator guidance
A credible open-weight alternative to closed frontier coding models at competitive hosted pricing — route here when self-hosting or open-weight licensing is a requirement, or when quality-per-dollar at long context matters more than picking the single highest-scoring closed model. The MiniMax Community License is not a fully permissive open-source license (see limitations) — confirm it fits the intended commercial use before relying on self-hosting.
Use cases
- Open-weight self-hosting where 1M-context, frontier-tier coding/agentic performance is needed without a closed-API dependency
- Long-document and long-conversation agentic workflows where MSA's context-length efficiency matters
Limitations
- Instruction-following and multilingual quality are not independently confirmed — omitted from capability_profile rather than guessed (no vendor or third-party benchmark found at time of writing)
- MiniMax Community License permits commercial use but requires a one-time notice to MiniMax under $20M annual revenue, or prior written authorization above it, plus mandatory "Built with MiniMax M3" attribution — not equivalent to a standard permissive license (Apache/MIT)
- Coding/reasoning ratings lean on vendor-published figures (MiniMax's own blog + Hugging Face model card) rather than independent third-party leaderboards — no independent SWE-bench Verified confirmation found at time of writing