MiniMax M3

M3 is a sparse mixture-of-experts language model with a vision encoder. The weights hold 427.0B parameters (the model card rounds this to ~428B), and about 23B are active per token. The published config lists 60 layers: the first three use dense feed-forward blocks, and the remaining 57 use 128 routed experts plus one shared expert, with four routed experts chosen per token by sigmoid scoring. The same 57 layers use MiniMax Sparse Attention, a block-sparse design built on grouped-query attention (64 query heads, 4 key-value heads); the first three keep dense attention. A lightweight index branch scores 128-token key-value blocks and keeps the top 16 for each group, and exact attention then runs only over those blocks. The MSA paper (arXiv 2606.13392) reports that on a 109B test model MSA matched GQA quality while cutting per-token attention compute 28.4x at 1M context. The vision side is a 32-layer ViT-style encoder with 14-pixel patches, images up to 2,016 pixels per side and 3D rotary positions for video. The vocabulary is 200,064 tokens, and the config carries multi-token-prediction modules for speculative decoding. MiniMax does not disclose training-token count, compute or knowledge cutoff.

Coding (8.8) and agentic work (8.8) are the design center. MiniMax reports SWE-bench Verified 80.5 and SWE-bench Pro 59.0, level with DeepSeek V4-Pro's 80.6 and ahead of GLM-5.1's 58.4 in MiniMax's own table. The strongest outside signal is the independent Long-Horizon Terminal-Bench: M3 scores 38.5, the best open-weights result on that board. Tool use backs up the agentic score (MCP Atlas 74.2, BrowseComp 83.5, OSWorld-Verified 75.2, all vendor-reported), and interleaved thinking between tool calls is a documented API feature. Vision (8.5) and document OCR (8.5) are real differentiators among open flagships: MMMU-Pro 78.1, Video-MME 85.4 and OmniDocBench 91.6, which tops every model in MiniMax's comparison table. Function calling (8.5) is strong, but M3 has no schema-constrained output mode: MiniMax's response_format works only with the older MiniMax-Text-01. Reasoning and math (8.0 each) rest on IMO 2025 at 35/42 and USAMO 2026 at 36/42. Those were produced with up to 10 test-time-scaling iterations and a 512K output budget, and MiniMax publishes no GPQA or HLE score. Long context (8.0) has the architecture for 1M, but the only long-context figure is LOCA-Bench at 256K (49.3 against Claude Opus 4.7's 57.0), and the API guarantees only 512K. Multilingual (8.0), instruction following (8.0) and creative writing (7.0) have no published benchmarks; LMArena's text ranking (1441, mid-table) is the only outside read on everyday chat quality. Real-time data (6.0) comes from a server-side web_search tool, in beta at $0.01 per request. Safety calibration (6.5) reflects a built-in content filter but no published safety framework.

Every row except the last two is vendor-reported from MiniMax's model card, and the harnesses matter. MiniMax ran SWE-bench and most coding rows on its own infrastructure with Claude Code as the scaffold and an overridden system prompt, averaged over four runs. It took the Terminal-Bench 2.1 scores for GPT-5.5, Gemini 3.1 Pro and Claude Opus 4.7 from the official leaderboard but ran M3 and the other models on its own infrastructure with a 128K output cap. The IMO and USAMO scores used up to 10 test-time-scaling iterations. The open-weights peers in MiniMax's table are older releases (GLM-5.1, Kimi K2.6), not the current GLM-5.3 or Kimi K3. The two independent results point in different directions. On Long-Horizon Terminal-Bench, a 46-task suite run with the Terminus-2 harness in a July 2026 snapshot, M3 is the top open-weights model, ahead of GLM-5.2 (31.6) and DeepSeek V4-Pro (30.7), though GLM-5.3 and Kimi K3 are not yet on that board. On LMArena's human-preference text leaderboard (vote cutoff 2026-09-13, 48,540 votes), M3 ranks 84th of 402 at 1441, well behind GLM-5.3, Kimi K3 and DeepSeek V4-Pro (1463). Artificial Analysis figures were not available to read (its site sits behind a bot checkpoint), so that column stays empty.

No independent throughput or time-to-first-token measurement of the MiniMax API was available, so both fields stay null. MiniMax's own docs quote about 100+ output tokens per second for M3. A Priority service tier (1.5x price) buys priority admission for faster responses and fewer failures under load. For self-hosters, the SGLang project publishes reference serving numbers on a 2,048-in / 256-out random workload at concurrency 64. On 8x B200 with MXFP8 weights it measured 1.58 s mean time to first token, 24.1 ms per output token and 2,385 tokens per second per GPU. On 8x H200 with BF16 weights it measured 1.05 s, 70.8 ms and 1,044 tokens per second per GPU. Thinking adds latency: on the OpenAI-compatible endpoint it is on unless disabled, so interactive apps should set thinking to disabled where reasoning isn't needed. The Long-Horizon Terminal-Bench run also shows M3 as one of the most persistent agents on that board, averaging 314 steps and about 90 minutes per task.

The price tier follows the size of each request's input. A request with 400K input tokens bills at $0.30 per 1M, while a 600K-token request bills at $0.60 per 1M, so its input costs $0.36 instead of $0.18. Workloads that routinely fill the window should budget for the higher tier. MiniMax's own prompt-caching page still works its example at the $0.60 list rate; the pay-as-you-go price list is the billing reference.

MiniMax API. The recommended path is the Anthropic-compatible endpoint (https://api.minimax.io/anthropic), which supports thinking blocks and interleaved thinking. OpenAI-compatible Chat Completions (https://api.minimax.io/v1) and an OpenAI Responses-compatible endpoint are also available. Images (up to 10 MB) and video (MP4, AVI, MOV or MKV, up to 50 MB inline or 512 MB through the Files API) go in as content blocks. The Anthropic endpoint ignores stop_sequences and top_k.

Third-party hosts. OpenRouter lists 13 providers, including Together, DeepInfra, Novita, Parasail, SambaNova and CoreWeave. They serve FP8 or FP4 weights, and max output ranges from 65,536 to 1,048,576 tokens depending on the host.

Self-hosting. BF16 weights (about 854 GB) and an official MXFP8 checkpoint (about 444 GB) are on Hugging Face and ModelScope, with recipes for SGLang, vLLM, Transformers, KTransformers, unsloth and AMD's ATOM (MXFP4/MXFP8). MiniMax's own self-hosting guide labels M3 support "Experimental" as of its 2026-08-26 review, still delivered through an SGLang development image. The reference configuration is 8x B200 with MXFP8 weights. SGLang also validates 4x B300 or GB300 (MXFP8), 8x H200 (BF16) and AMD MI355X/MI300X for text. SGLang reports validation only from 1K to 128K input tokens, and video serving is not validated in that recipe. Peak GPU memory is not published, so min_gpu_vram_gb stays null.

License. Self-hosting is where the license bites. The MiniMax Community License grants rights to use, copy, modify and distribute the weights "for non-commercial purposes". Any commercial use, including products built on derivatives, has three conditions:

• The product must prominently display "Built with MiniMax M3" on a related website, user interface, blog post, about page or product documentation.
• Products or services with more than US$20 million in yearly revenue need separate prior written authorization from MiniMax (api@minimax.io, subject "M3 licensing - authorization request").
• Everyone else must send MiniMax a one-time notice (subject "M3 licensing — notice").

"Commercial Use" is defined broadly. It covers paid products that rely on the model, commercial use of APIs serving the model, and commercial deployment of fine-tuned or otherwise modified versions. Counsel should confirm whether that reaches a product built on MiniMax's own API or on a third-party host's M3 endpoint. An appendix also prohibits, among other things, any military purpose. This is materially tighter than the licenses of the peers named below, all of which grant commercial use up front and add conditions only at scale.

MiniMax publishes no safety framework, system card or refusal-calibration data for M3, and lists no compliance certifications (SOC 2, ISO 27001, HIPAA) for the API. The first-party API runs sensitive-content detection on every request. Responses carry input_sensitive and output_sensitive flags with a category code (pornography, violence or terrorism, abuse, prohibited content and others), and a severe input violation returns an empty reply. The API privacy policy (effective 2026-03-30) names Nanonoble Pte. Ltd. in Singapore as the operator. It says personal data is stored in a US data center and processed under GDPR-style safeguards, with jurisdiction-specific rights for the EEA, UK and several US states. It lists "messages and content" among data used to analyze and improve the services, and it does not describe an opt-out for model training on API inputs, so trains_on_inputs stays null. Self-hosting removes the data-flow question but not the license's prohibited-use list, which bars military use, harm to minors and harmful disinformation.

MiniMax serves M3 through Anthropic-compatible, OpenAI-compatible and Responses-compatible APIs, and documents the Anthropic SDK, OpenAI SDK and Vercel AI SDK in Python and TypeScript. That makes it a drop-in backend for Claude Code, Cursor and other tools that accept custom endpoints, packaged commercially as the Token Plan ($22 to $132 a month). MiniMax's own MiniMax Code agent platform runs on M3. Self-hosting has day-one recipes for SGLang, vLLM, Transformers, KTransformers, unsloth and AMD ATOM, plus an official MXFP8 checkpoint and ModelScope mirrors. Adoption is growing. The Hugging Face repo shows about 165,000 downloads in the last month and 1,500 likes, 13 providers serve it on OpenRouter, and LMArena has collected more than 48,000 votes on it. It is still smaller than the GLM and Kimi communities.