MiniMax

AI Coding Agents Report: August 2026 · Updated 02 September 2026 · Version history

Executive Summary

What it is: MiniMax is a Chinese frontier lab selling its M-series models through platform.minimax.io (global), platform.minimaxi.com (China), Alibaba's Bailian marketplace, and the Hailuo AI apps. The coding-relevant product is MiniMax-M3, a 428B-parameter MoE (~23B active) described by MiniMax as its "frontier multimodal coding model with 1M context window," served at ~100+ tokens/sec via recommended Anthropic-compatible endpoints (api.minimax.io/anthropic), with a Token Plan subscription for using M3 inside Claude Code and Cursor. On Bailian, M3 lists at ¥4.2/¥16.8 per MTok. Source: https://platform.minimax.io/docs/guides/text-generation , https://help.aliyun.com/zh/model-studio/billing-for-model-studio

What to watch out for: MiniMax publishes no dedicated coding-agent harness (adoption flows through Claude Code, Cursor, and third-party platforms), its own Token Plan prices are not published on the public docs fetched this cycle, and the August open-weight milestone, MiniMax-H3, is a video-generation system whose Community License requires a separate application for USA/EU/UK/South Korea use, not a coding model. Enterprise features (SSO, audit, indemnity, SLA) are absent. Source: https://huggingface.co/MiniMaxAI/MiniMax-H3

Bottom line: MiniMax earns tracking through M3, a credible open-ecosystem coding model with 1M context, Claude Code/Cursor integration docs, and Bailian marketplace distribution, in the same supplier class as Meituan and DeepSeek; treat it as a model-and-API supplier, not a managed agent platform, until it ships a first-party harness. Source: https://platform.minimax.io/docs/guides/text-generation

Key Terms

  • MiniMax-M3 - flagship M-series model (June 1 on Bailian): 428B total / ~23B active MoE, 1M-token context, multimodal input (text, image, video), positioned for agentic reasoning, tool use, coding, and structured task execution at ~100+ tps output. Source: Minimax – Text Generation , Aliyun – Newly Released Models
  • Anthropic-compatible endpoint - https://api.minimax.io/anthropic is MiniMax's recommended API path, supporting thinking blocks and interleaved thinking, which makes M3 a drop-in backend for Claude Code and Anthropic-SDK tools; an OpenAI-compatible /v1 path also exists. Source: Minimax – Text Generation
  • Token Plan - MiniMax's subscription for using models "of all modalities at ultra-low prices," including documented integration for using M3 in Claude Code, Cursor, and other tools. Prices live on the subscribe page, not the public docs. Source: Minimax – Text Generation
  • MiniMax-H3 - 33B dense omni-modal generative system (open weights August 3) that understands text/image/video/audio and generates 4-15 second video with native stereo audio up to 2K. Not a coding model; it is the first fully open omni-modal model. Released under the MiniMax H3 Community License with a separate application form for USA/EU/UK/South Korea. Source: Hugging Face – Minimax H3
  • Highspeed variants - M2.7/M2.5/M2.1-highspeed deliver the same performance at ~100 tps versus ~60 tps standard. Source: Minimax – Text Generation

Latest Changes

First report for this supplier (added mid-cycle, September 2, after August's [NEW ENTRANT] flag in Market Dynamics). All models, plans, and pricing are listed as current state.

  • [NEW ENTRANT] added to tracking: MiniMax enters the tracked list on the strength of M3's coding positioning and marketplace distribution. August's trigger event was the H3 open-weight release (below); the coding case rests on M3. Source: Local-Ai-Zone – Ai Updates August 2026
  • Open-weight milestone (Aug 3): MiniMax-H3 weights published on HuggingFace (33B, BF16, FL2VA and Ref2VA checkpoints) with SGLang, vLLM, diffusers, and ComfyUI deployment recipes. 5,532,597 downloads in the last month, 4.77k likes, 59 community quantizations, 109 finetunes. Source: Hugging Face – Minimax H3
  • Marketplace distribution (Aug 17): MiniMax/MiniMax-H3 listed on Alibaba Bailian, where the text family (M3, M2.7, M2.5, M2.1) was already served. Source: Aliyun – Newly Released Models
  • Coding integration: MiniMax's docs include a dedicated guide for running M3 inside Claude Code and Cursor via the Token Plan, alongside Anthropic/OpenAI-compatible endpoints. Source: Minimax – Openclaw

Plans

Plan Price Notes
Pay-as-you-go API Per-token Via platform.minimax.io (global) / platform.minimaxi.com (CN)
Token Plan See subscribe page (not on public docs) "Ultra-low prices" across all modalities; required surface for Claude Code/Cursor M3 use. Source: Minimax – Token Plan

No consumer coding-agent subscription; Hailuo AI apps cover chat/video. Source: Minimax – Text Generation

API Pricing

Model (Bailian, Beijing) Input ¥/MTok Output ¥/MTok Context Notes
MiniMax/MiniMax-M3 ¥4.2 (~$0.58) ¥16.8 (~$2.35) 1M Thinking and non-thinking; context caching discounts
MiniMax/MiniMax-M2.7 ¥2.1 ¥8.4 204,800 Thinking only
MiniMax/MiniMax-M2.5 ¥2.1 ¥8.4 204,800 Thinking only
MiniMax/MiniMax-M2.1 ¥2.1 ¥8.4 204,800 Thinking only

MiniMax's own platform rates were not published on the public docs pages fetched this cycle; the subscribe page carries Token Plan pricing. Source: Aliyun – Billing For Model Studio , Minimax – Text Generation

Model Performance / Benchmarks

Benchmark Score Notes
GPQA 92.9% M3, per aggregator coverage of the June launch. Source: Local-Ai-Zone – Ai Updates August 2026
Coding/agent benchmarks Not published No MiniMax-published SWE-Bench/Terminal-Bench table found this cycle.

Latest News

H3: First Fully Open Omni-Modal Model (August 3, 2026)

MiniMax published full weights for H3, a 33B omni-modal system generating 2K video with native stereo audio, with production-grade deployment recipes and reproducible test cases. It is a generative-media milestone rather than a coding event, but it anchors MiniMax's open-ecosystem credibility (5.5M downloads in a month). License caveat: Community License with a required application for USA/EU/UK/South Korea. Source: Hugging Face – Minimax H3

M3 Rides the Marketplace Wave (August 2026)

With M3 on Bailian since June and H3 added August 17, MiniMax models are now consumable through the same Chinese marketplace hosting GLM-5.3, Kimi K3, and DeepSeek V4, and through Western coding tools via Anthropic-compatible endpoints. Source: Aliyun – Newly Released Models

Community Signals

The H3 HuggingFace page shows exceptional ecosystem uptake for a first release: 5,532,597 downloads last month, 102 discussions, 59 quantizations, 109 finetunes, and 100+ demo Spaces including acceleration arenas and inpainting experiments. Source: Hugging Face – Minimax H3 . No significant HN threads on M3 as a coding backend were found this cycle.

Enterprise Readiness

Feature Available? Details
SSO (SAML/OIDC) No Not documented.
SCIM No Not documented.
Audit logs No Not documented.
IP indemnity No Not documented.
Data residency Partial Distinct global (minimax.io) and China (minimaxi.com) platforms; no residency commitments published. Source: Minimax – Text Generation
HIPAA No Not advertised.
Air-gapped / on-prem Partial H3 weights are downloadable (video system); M-series weights are not open, so coding workloads cannot be self-hosted. Source: Hugging Face – Minimax H3
SLA No Not published.
Admin controls (RBAC) No Not documented.

Transparency Gaps

Gap Details Severity
Platform pricing MiniMax's own per-token rates and Token Plan prices are not on the public docs pages; Bailian rates are the only verifiable numbers. High
Coding benchmarks No published SWE-Bench/Terminal-Bench/CursorBench scores for M3 despite "frontier coding model" positioning. High
H3 license scope The Community License's restrictions and the USA/EU/UK/KR application criteria are not summarized in-machine-readable form; commercial status differs by region. Medium
Harness roadmap Whether MiniMax will ship a first-party coding agent (like Zhipu's ZCode) is unknown. Low

← Back to AI Coding Agents Report: August 2026