MiniMax-M3
v20260601MiniMax
MiniMax's natively multimodal 428B MoE (23B active) with a 1M-token context via MiniMax Sparse Attention, launched June 2026 with weights on Hugging Face by 2026-06-07. Vendor reports 80.5% SWE-bench Verified; Artificial Analysis scores it 44 on Intelligence Index v4.1. Unlike MIT-licensed M2, M3 ships under a MiniMax Community License, and training code and some inference operators are withheld.
Trust Vector Analysis
Dimension Breakdown
🚀Performance & Reliability+
Meaningful step over M2: 1M context, native multimodality, and independently confirmed intelligence gains (Artificial Analysis 44 vs open-weight median 25). Headline coding figures (80.5% SWE-bench Verified) are vendor-run on vendor infrastructure and drew explicit independent skepticism at launch; GLM-5.2 leads it on independent indices.
Vendor benchmarks with independent scrutiny noting all launch figures were vendor-run; partial independent corroboration via Artificial Analysis
Vendor-reported reasoning benchmarks and early community evaluation
Independent composite benchmarking (Artificial Analysis) plus vendor multimodal benchmarks
Early community testing of repeated runs and agentic trajectories
Median latency for API requests with standard prompt sizes; architecture-level efficiency verified in vendor technical report (arXiv:2606.13392)
95th percentile response time from early third-party measurements
Official specification from model card
Review of platform availability since launch; observation window under six weeks
🛡️Security+
Standard open-model posture without third-party audits, scored slightly below M2 for launch recency and the wider multimodal attack surface. Self-hosting shifts security responsibility to the deployer.
Review of vendor documentation against OWASP LLM01 patterns; model too new for mature red-team coverage
Early testing against adversarial prompt datasets; deployer-dependent for self-hosted use
Analysis of privacy policies and self-hosting data-control options
Safety testing across harmful content categories per vendor card and early community reports
Review of API security features and best practices
🔒Privacy & Compliance+
First-party MiniMax API operates under Chinese jurisdiction — a material caveat for Western regulated industries. Self-hosting mitigates residency, but unlike MIT-licensed M2, M3's Community License attaches attribution and revenue-threshold conditions to commercial self-hosted use.
Review of provider jurisdiction and third-party hosting options
Analysis of privacy policy and data usage terms
Review of terms of service and deployment-dependent retention
Review of data protection capabilities and customer responsibilities
Verification of compliance certifications and audit reports
Review of self-hosting deployment options enabling zero retention
👁️Trust & Transparency+
Good architectural documentation (MSA technical report) but reduced openness vs M2: training code, data pipelines, and some inference operators are withheld, and all launch benchmarks were vendor-run — independent reviewers explicitly flagged the verification gap.
Evaluation of reasoning transparency and trajectory inspectability
Early testing on factual QA and tool-augmented workflows
Review of published bias benchmarks and community evaluations
Qualitative assessment of confidence expression in outputs
Review of documentation completeness and clarity
Review of public disclosures about training data and released artifacts
Analysis of built-in safety mechanisms
⚙️Operational Excellence+
Dual OpenAI/Anthropic compatibility and broad day-one inference support carry over from M2, but the license regression (MIT to MiniMax Community License with attribution and revenue-threshold clauses) and withheld training/inference code lower the operational score. Commercial deployers need legal review.
Review of API design, consistency, and feature completeness
Review of SDK quality, documentation, and maintenance
Review of versioning practices, weight availability, and license continuity across releases
Review of available monitoring tools and metrics
Assessment of documentation, community, and support responsiveness
Analysis of third-party hosting, integrations, and tooling; conservative given launch recency
Review of licensing terms and restrictions; regression from M2's clean MIT license is trust-relevant
- +Native multimodality: mixed-modality training from step one across text, image, and video (78.1 MMMU Pro, 85.4 Video-MME v2)
- +1M-token context with MiniMax Sparse Attention: 9x prefill / 15x decode speedups vs M2 at 1M context
- +Efficient 23B-active footprint keeps inference fast and self-hosting relatively affordable
- +Aggressive pricing: $0.30/$1.20 per 1M tokens (promotional 50% off list) with $0.06 cache reads
- +Independently confirmed intelligence gain: Artificial Analysis 44, well above the open-weight median
- +Day-one vLLM/SGLang/Transformers/KTransformers/unsloth support with community quantizations
- !License regression from M2: MiniMax Community License requires 'Built with MiniMax M3' attribution and separate authorization above a revenue threshold
- !Training code, data pipelines, and some inference operators withheld despite the 'open-weight' label
- !All launch benchmarks vendor-run on vendor infrastructure; independent reviewers flagged the verification gap
- !First-party API processes data under Chinese jurisdiction with no published Western compliance certifications
- !Behind GLM-5.2 on independent intelligence and agentic indices
- !Long-context pricing doubles above 512K input tokens
- !Launched June 2026 — consistency, uptime, and security evidence still immature
Use Case Ratings
code generation
Vendor-reported 80.5% SWE-bench Verified and 59.0% SWE-bench Pro at very low cost, but figures are vendor-run and GLM-5.2 leads on independent indices.
customer support
Fast (23B active), very cheap, and multimodal — can handle screenshot- and image-based support tickets natively.
content creation
Good generation quality at minimal cost; image/video understanding aids content workflows.
data analysis
Good tool calling plus native chart/image understanding; weaker raw reasoning than GLM-5.2 or Kimi K2.6.
research assistant
1M context and native multimodality (text/image/video) suit long-document and mixed-media research at low cost.
legal compliance
China-jurisdiction first-party API, no Western certifications, and Community License conditions make this a poor fit without careful mitigation.
healthcare
Not recommended via first-party API; self-hosted deployment in a compliant environment is the only viable path, with license review.
financial analysis
Capable and cheap with chart/document understanding; data residency requires self-hosting for regulated firms.
education
Multimodal tutoring (diagrams, video) at very low cost suits high-volume educational platforms.
creative writing
Serviceable creative output; not its design focus.