MiniMax M3
- Provider
- MiniMax
- Status
- available
- Context
- 1,000,000 tok
- SWE-bench
- 80.5%
- Price
- $0.3 / $1.2 /MTok
MiniMax M3 is MiniMax’s open-weight flagship, released on 1 June 2026 with frontier-level coding, a 1-million-token context window and native multimodality — all in one open-weight model. Its headline is value: MiniMax M3 beats GPT-5.5 and Gemini 3.1 Pro on key benchmarks at roughly 5–10% of their cost (VentureBeat), and it is downloadable and self-hostable rather than API-only.
On coding it is genuinely frontier-adjacent: 59.0% on SWE-bench Pro — above GPT-5.5 and Gemini 3.1 Pro, approaching Claude Opus 4.7 — and around 80.5% on SWE-bench Verified, within ~0.2 points of Gemini 3.1 Pro (OpenRouter). It sits in the open-weight cluster with DeepSeek V4 and Kimi K2.6 that has closed much of the gap to the proprietary leaders.
Quick specs
| Provider | MiniMax |
| Released | 1 June 2026 |
| Status | Available (open weights) |
| Architecture | Mixture-of-experts, natively multimodal |
| Context window | 1,000,000 tokens |
| Licence | Open weights (self-hostable) |
| Input price | ~$0.30 / MTok |
| Output price | ~$1.20 / MTok |
| SWE-bench Pro | 59.0% (beats GPT-5.5, Gemini 3.1 Pro) |
| SWE-bench Verified | ~80.5% |
| Best for | Low-cost / self-hosted frontier-adjacent coding and agentic work |
| Limitations | Below the proprietary frontier on hardest tasks; China-hosted |
What MiniMax M3 is
MiniMax M3 is a natively multimodal, mixture-of-experts model built for coding and agentic work, with a 1M-token context window drawing on MiniMax’s heritage of efficient long-context architectures (Pandaily). The point of the release is that it delivers proprietary-class results open and cheap: the API launched first on 1 June 2026, with the open weights released shortly after for full enterprise download and customisation.
That combination — frontier-adjacent quality, open weights, and pricing at roughly a tenth of US labs — is what made M3 notable, and why it features in our best AI models ranking’s open-weight tier.
Benchmark performance
MiniMax-reported at launch, with independent context; treat vendor numbers as a ceiling.
| Benchmark | MiniMax M3 | Notes |
|---|---|---|
| SWE-bench Pro | 59.0% | Above GPT-5.5 and Gemini 3.1 Pro; approaches Claude Opus 4.7 |
| SWE-bench Verified | ~80.5% | Within ~0.2 points of Gemini 3.1 Pro |
| Terminal-Bench 2.1 | 66.0% | Strong agentic terminal coding |
| BrowseComp | 83.5 | Web-browsing/agentic |
| MCP-Atlas | 74.2% | Tool use |
The pattern is strong, cost-efficient coding and agentic performance (VentureBeat). It sits below the closed frontier (Claude Opus 4.8, GPT-5.5) on the very hardest tasks and on long-horizon agentic reliability, but for the price it is exceptional. See best AI for coding for cross-model standings.
Pricing and access
MiniMax M3 is priced at roughly $0.30 input / $1.20 output per million tokens via MiniMax’s API platform — about 5–10% of the cost of GPT-5.5 or Gemini 3.1 Pro for comparable benchmark results (VentureBeat). The open weights are free to download and self-host (typically via Hugging Face), enabling air-gapped, data-sovereign deployment.
How MiniMax M3 compares
- vs the proprietary frontier — M3 beats GPT-5.5 and Gemini 3.1 Pro on several coding benchmarks and approaches Claude Opus 4.7, at a tiny fraction of the price; the closed leaders (Claude Opus 4.8, GPT-5.5) still lead on the hardest tasks and long-horizon reliability.
- vs other open models — In the open-weight cluster with DeepSeek V4, Kimi K2.6 and Alibaba’s Qwen, M3 competes on coding, its 1M context, native multimodality and very low price.
Known limitations
Below the closed frontier on the hardest reasoning and most demanding long-horizon agentic work. China-hosted — the hosted API raises data-residency and compliance questions for Western enterprises and governments (open weights mitigate this via self-hosting). Vendor benchmarks are a ceiling — prefer standardized leaderboards where decisions ride on the number.
FAQ
What is MiniMax M3?
MiniMax M3 is MiniMax’s open-weight flagship, released 1 June 2026 — a natively multimodal, mixture-of-experts model with a 1M-token context, focused on coding and agentic work, that matches or beats GPT-5.5 and Gemini 3.1 Pro on key benchmarks at a fraction of the cost.
How much does MiniMax M3 cost?
About $0.30 per million input tokens and $1.20 per million output via MiniMax’s API — roughly 5–10% of the cost of comparable proprietary models. The open weights are free to self-host.
Is MiniMax M3 open source?
It is released as open weights, free to download and self-host (with the API launching first and weights following shortly after). Check the model card for the specific licence terms.
How good is MiniMax M3 at coding?
Very strong for the price: 59.0% on SWE-bench Pro — ahead of GPT-5.5 and Gemini 3.1 Pro and approaching Claude Opus 4.7 — and around 80.5% on SWE-bench Verified, within ~0.2 points of Gemini 3.1 Pro.
Is MiniMax M3 better than DeepSeek V4?
They are close peers in the open-weight tier. DeepSeek V4 leads marginally on SWE-bench Verified (80.6% vs 80.5%) under an MIT licence; MiniMax M3 counters with stronger SWE-bench Pro, native multimodality and a 1M context. For self-hosted, low-cost coding both are excellent choices.
Last verified 18 June 2026. MiniMax M3 figures are MiniMax-reported from the 1 June 2026 launch, with independent context from VentureBeat and OpenRouter; pricing and licence terms should be confirmed on MiniMax’s platform and the model card before relying on them.