MiniMax-M3

What it is. MiniMax-M3 (2026-06-01) is MiniMax’s flagship model: a ~428B-total / ~23B-activated MoE with MiniMax Sparse Attention, a 1M-token context window, and native multimodal input. Why it matters. It targets agentic reasoning, tool use, coding and long-context work at one of the lowest flagship prices in the database ($0.30 input). Key characteristics. Text, image (up to 10MB) and video (up to 50MB direct, 512MB via Files API) input with text output; thinking disabled by default with a thinking=adaptive toggle. What a professional should know. Maximum output tokens are not publicly disclosed — the 131,072 figure is derived from the official card’s evaluation config, not a published API ceiling — and pricing doubles above the 512K input boundary.

Architecture and parameters

MiniMax-M3 is a ~428B-total / ~23B-activated MoE (~5.4% activation) built around MiniMax Sparse Attention (MSA), which MiniMax claims delivers 9x prefill and 15x decode speedup vs M2 at 1M context. That speedup is a vendor claim, not an independent measurement, but the architectural direction — linear/sparse attention to make long context affordable — matches the market-wide pattern (Qwen’s DeltaNet, Kimi’s KDA). See Mixture-of-Experts and the MoE architectures research.

What the context window actually means

The 1M-token window comes with a practical qualifier: at least 512K is guaranteed usable, and pricing splits at the 512K boundary (double above). This is honest tiering — the full 1M is advertised, but the economics and the guarantee both live at 512K. Treat M3 as a 512K-usable, 1M-capable model. See the context-window research.

Pricing implications

International pay-as-you-go standard tier is $0.30 input / $1.20 output per 1M tokens up to 512K input (double above 512K), cache reads $0.06, and a permanent 50% off list — the priority tier costs 1.5x. China AI Hub analysis: at $0.30, M3 is the cheapest flagship input price in the database, which is its clearest market position. The 4x input-to-output ratio is standard. See the pricing-changed research.

API, coding, and agent implications

Vendor-reported SWE-bench Pro (59.0), Terminal-Bench 2.1 (66.0), MCP Atlas (74.2), BrowseComp (83.5) and PostTrainBench (37.1) describe a capable agentic/coding profile with particular strength in MCP-based tool use and browsing. Agent capability is documented, and tool calling plus multimodal input make it a strong agent substrate. Thinking is off by default — unlike GLM-5.3’s always-on — so you opt into reasoning per request via thinking=adaptive. See the Doubao comparison and Kimi comparison.

Open weights and license

Open weight (MXFP8, ~171k downloads) on Hugging Face/GitHub under the MiniMax Community License — a custom license that is free for non-commercial use but requires prominent “Built with MiniMax M3” attribution for commercial use, and written authorization above $20M yearly revenue. This is conditional-open, not permissive. See licensing explained.

Benchmark interpretation

All five rows are vendor-reported. BrowseComp 83.5 and MCP Atlas 74.2 are agentic/browsing/tool-use signals; SWE-bench Pro 59.0 and Terminal-Bench 66.0 are coding; PostTrainBench 37.1 is a post-training quality metric. These describe a mid-to-strong coding profile at a low price, not a frontier-reasoning leader.

What the benchmarks do not prove

The scores are vendor-reported and not independently verified; the MSA speedup claims (9x/15x) are unmeasured by any third party; and the maximum-output ceiling is unpublished. These gaps matter when M3’s headline advantages — speed and price — are exactly what you’d want independently confirmed. See how to read vendor-reported benchmarks.

Suitable and less suitable workloads

Well-suited: cost-sensitive agentic work, MCP/tool-use applications, multimodal input (image/video) analysis, browsing agents, and long-context tasks within the 512K guaranteed band. Less suited: frontier-level deep reasoning, workloads requiring a documented output ceiling, and commercial self-hosting above the license thresholds.

Where this model fits

Workload	Relevance
Long-context analysis	High
Coding	High
Structured API workflows	Moderate
Agent orchestration	High
Local self-hosted deployment	High
GUI automation	No evidence
Video generation	No evidence
Enterprise cloud	Moderate

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

China AI Hub analysis

China AI Hub analysis indicates M3’s competitive edge is price-per-capability in the agentic tier: it delivers multimodal input, MCP tool use and a 1M window at the lowest flagship input price tracked. The trade-offs are verification depth (vendor-only benchmarks, unpublished max output, unmeasured speed claims) and the 512K pricing split, which means the advertised 1M context costs 2x in the upper band. For budget-constrained agent deployments it is a strong candidate; for workloads where a hard, documented output ceiling or independently verified performance is required, the gaps are real.

Market position and outlook

China AI Hub analysis indicates M3’s edge is price-per-capability in the agentic tier: it delivers multimodal input (text plus JPEG/PNG/GIF/WEBP images up to 10MB and video up to 50MB direct or 512MB via the Files API), MCP tool use and a 1M window at the lowest flagship input price tracked ($0.30). The MCP Atlas score (74.2) is worth singling out because it measures the model’s ability to drive MCP-connected tools, which is exactly the integration pattern agent builders use. The trade-offs are verification depth — vendor-only benchmarks, an unpublished maximum-output ceiling, and unmeasured 9x/15x speed claims — plus the 512K pricing split that doubles cost in the upper context band. For budget-constrained agent and multimodal deployments it is a strong candidate; for workloads needing a hard documented output ceiling or independent performance evidence, those gaps remain. See the pricing-changed and multimodal capabilities research.

What is uncertain

• Maximum output tokens are not publicly disclosed; the 131,072 figure is derived from the official card’s evaluation config.
• The MSA speedup claims (9x prefill / 15x decode) are vendor claims, not independently measured.
• Benchmark scores are vendor-reported and not independently verified.

Sources

• MiniMax API platform — model overview (CN)
• MiniMax official M3 model page
• MiniMax M3 official blog post
• Hugging Face model card — MiniMax-M3

Labels used above: Official fact (from primary sources), Vendor-reported claim (benchmark scores and MSA speedup claims published by MiniMax), and China AI Hub analysis (our synthesis, always introduced as such). No third-party benchmark evidence is currently recorded for this model.