Skip to main content
Chinese AI Tools
ProductsModelsIntegrationsRankingsLatest changes
TopicsAvailability TrackerUse casesSubmit a toolAccount
ZHSearch

Chinese AI Tools

Independent directory for Chinese AI products. Product availability, pricing and terms can change. Verify before commercial use.

Editorial standardsClaim productUpdate infoGet featuredAdvertise

MiniMax

MiniMax M3

MiniMax's frontier coding and agentic model with 1M context and native multimodal understanding.

Globally availableFull English UIPublic APIFreemium

Quick answer

M3 should be tracked separately because the official homepage now gives it a distinct model page and positions it as the current frontier coding model.

Official site
Pricing
API & Token Plan access; current pricing should be checked in account
Availability
Globally available
API
Public API
Last checked
2026-09-07
Pricing detailsUse-case fitSource evidence

MiniMax M3 was released on June 1, 2026 as MiniMax's flagship for coding, agents and native image/video understanding. Its MiniMax Sparse Attention architecture supports up to a 1M-token API context window with a guaranteed minimum of 512K. MiniMax reports that at 1M context its per-token compute is 1/20 of the previous generation, with more than 9x faster prefill and 15x faster decoding. The API uses the MiniMax-M3 model identifier, supports thinking and non-thinking modes, applies a higher long-context rate above 512K input tokens, and offers an optional priority service tier for SLA-sensitive workloads. The official 427B-parameter weights are now live on Hugging Face under the MiniMax Community License, even though the checked product page still contains older coming-soon copy.

Topic guide

MiniMax models, agents and creative platform

A source-backed map of MiniMax M3, API and Token Plan access, Mavis Agent, MiniMax Design, Hailuo, speech, music and image products.

Editorial verdict

Best for

Developers evaluating a China-origin frontier model for coding, long-context agentic work and multimodal reasoning.

Avoid if

Avoid it if you need fixed public pricing or a model with fully stable commercial terms already locked in.

Why it matters

M3 should be tracked separately because the official homepage now gives it a distinct model page and positions it as the current frontier coding model.

Trust: 4/4 sources verified, recently checkedCoverage: 100/100

Pricing

API & Token Plan access; current pricing should be checked in account

Payment

API & Token Plan, Pay-as-you-go API billing, Enterprise sales

Commercial use

Commercial use should follow the current product, API, model license and billing terms.

Privacy

Review prompt, file, media upload, retention and training-use terms before sensitive workloads.

Use-case fit

Frontier coding evaluation

Strong

Evaluate repository repair, terminal execution, kernel optimization and MCP tool use, while reproducing vendor benchmark settings independently.

Long-context agent workflows

Strong

The 1M context window makes it relevant for large codebases, long documents and multi-step agent tasks.

Multimodal reasoning

Strong

Use native image and video input for document understanding, visual coding tasks and computer-use agents.

Long-horizon autonomous work

Medium

The release reports 12-hour paper reproduction and 24-hour CUDA optimization runs; validate cost, reliability and recovery behavior on your own tasks.

Global user checklist

RegistrationPartialThe homepage exposes API and Token Plan flows, but the exact account path should be rechecked in the target region.
English UIConfirmedThe homepage, model page and report page are English-facing.
API and docsConfirmedDocs expose the MiniMax-M3 identifier and chatcompletion_v2 endpoint, up to 1M context with a guaranteed 512K minimum, thinking-mode selection, the 512K pricing boundary and standard versus priority service tiers.
International paymentPartialAPI and Token Plan access are public, but live prices and quotas should supersede the release-article snapshot.
Commercial usePartialProduction users should verify model-weight licensing, API terms, priority-tier access and multimodal data handling.

Rechecked the official product and weight pages on September 7, 2026. The official Hugging Face repository hosts the weights under the MiniMax Community License, while the product page still contains older coming-soon copy. Recheck current API prices, quotas and priority-tier availability before rollout.

Pros

  • - Vendor-reported scores include 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1 and 74.2% on MCP Atlas
  • - API context is documented at up to 1M tokens with a guaranteed minimum of 512K
  • - Native image and video understanding includes computer-use scenarios
  • - Thinking and non-thinking modes can be selected per API request
  • - Vendor reports an 83.5 BrowseComp score and long-horizon autonomous runs lasting 12 to 24 hours

Cons

  • - Most comparative scores and real-world demonstrations are vendor-run and need independent reproduction
  • - Inputs above 512K tokens use a higher long-context API rate
  • - The priority service tier requires sales access at the time of the release article
  • - The MiniMax Community License and 427B-parameter serving requirements need review before commercial self-hosting

Decision paths

minimax-api

qwen

deepseek

zhipu-glm

kimi

Sources

MiniMax M3 model page

Official · EN · verified 2026-09-07

Confirms coding and agent positioning, guaranteed 512K and maximum 1M API context, native multimodality, BrowseComp score and long-horizon demonstrations. Its weight-publication copy is older than the live official repository.

MiniMax M3 report

Official · EN · verified 2026-06-10

Documents the June 1 release, MSA efficiency, benchmark methodology, real-world agent runs, API modes and access routes.

MiniMax M3 official model weights

Repository · EN · verified 2026-09-07

Confirms that the 427B-parameter weights are live under the MiniMax Community License, with about 23B active parameters, 1M context and text, image and video inputs.

MiniMax homepage

Official · EN · verified 2026-06-01

Shows M3 alongside Hailuo 2.3, Speech 2.8 and Music 2.6 on the public homepage.

Last checked: 2026-09-07

Reviews

No approved reviews yet.

Availability snapshot

Availability
Globally available
English UI
Full English UI
API
Public API
Rating
4.6 (0)

Latest updates

Latest changes
Release · 2026-08-25

Qwen, Kimi and MiniMax ecosystem topics published

Three new topic hubs connect each ecosystem's current models, coding agents, API or subscription paths, multimodal products and practical selection guides. Ownership and product boundaries are called out where similarly named services come from different teams.

Pricing · 2026-08-04

MiniMax multimodal API pricing refreshed

MiniMax's live API pricing lists permanently discounted M3 rates for context up to 512K at $0.30/M input, $1.20/M output and $0.06/M cache-read tokens; 512K-1M costs $0.60, $2.40 and $0.12, with access above 512K currently limited. M2.7 costs $0.30/M multimodal input and $1.20/M output, while Highspeed costs $0.60 and $2.40. Hailuo 2.3 starts at $0.28 per 768P six-second video, Speech 2.8 Turbo at $60/M characters, rapid voice cloning at $1.50 per voice, voice design at $3 per voice, and Music 2.6 at $0.15 per song with a free-trial label.

Release · 2026-06-30

OpenRouter Owl Alpha added

OpenRouter's Owl Alpha is now tracked as a free text model for agentic coding, tool use, automated workflows and complex instruction execution. The live model API lists a 1,048,756-token context window, up to 262,144 output tokens, native tools and structured outputs, zero input and output token prices, and one Stealth INT8 provider route. OpenRouter says it works with Claude Code and OpenClaw, but also warns that the provider may log prompts and completions and use them for model improvement. The underlying model identity and license remain undisclosed, so the profile treats it as a non-sensitive Alpha evaluation route rather than a production-safe default.

Open source · 2026-06-26

DeepReinforce Ornith 1.0 open coding models added

DeepReinforce's Hugging Face organization now exposes the Ornith 1.0 family of MIT-licensed agentic-coding models. The collection includes public 9B, 35B and 397B model cards plus GGUF and FP8 variants. The cards position Ornith as a self-improving model family post-trained from Gemma 4 and Qwen 3.5, using reinforcement learning to optimize both solution rollouts and their scaffolds, with reported results on Terminal-Bench 2.1, SWE-Bench, NL2Repo, SWE Atlas and ClawEval. The profile records vLLM, SGLang, Transformers and OpenAI-compatible local serving paths.

Submit a review