All AI models
Open-weightVerified October 23, 2026

MiniMax

MiniMax M3

MiniMax M3 is a strong open-weight candidate for teams that value a long context window, native multimodal input, agentic coding, and the ability to decide how much reasoning a request deserves. Start on the API, then evaluate self-hosting only if data control or sustained scale justifies the operational work.

What changed: Newer from MiniMax: MiniMax M3.1 Flash, released October 19, 2026 as an open-weight model, costs $0.15 input and $0.60 output per 1M tokens. Its size and license were not stated consistently on vendor pages as of October 23, 2026.
License
MiniMax community license
Context
1M tokens
Modalities
Text · Image · Video input
Parameters
About 428B total, about 23B active

What the license means

MiniMax publishes the M3 weights, and the Hugging Face model card lists the license as "minimax-community", a custom license rather than Apache or MIT. Read those terms before redistribution, offering it as a service, or building a commercial derivative.

Open-weight, open source, and no-cost API access mean different things. Always review the linked terms and your own data, hosting, and compliance requirements before shipping.

API and operating cost

Input / 1M
$0.30
Output / 1M
$1.20

Standard tier for inputs up to 512K tokens, after a discount MiniMax labels permanent (list price $0.60 / $2.40). Inputs over 512K cost $0.60 / $2.40 after the discount.

Check current provider price

Where it fits

  • Agentic coding
  • Large codebases and document sets
  • Image and video-aware automation

Where it does not fit

  • Small local machines
  • Simple chat where 1M context adds no value
  • Teams that have not reviewed the current model-card terms

Local deployment reality

MiniMax documents SGLang, vLLM, Transformers, KTransformers, and Unsloth. Its 428B total parameter footprint makes managed access the practical starting point for most builders.

SGLangvLLMTransformersKTransformersUnsloth

Practical setup notes

  • Use adaptive reasoning before defaulting every request to the most expensive thinking setting.
  • Keep static instructions at the beginning of a prompt to improve cache reuse.

Capability notes

Reasoning
Enabled, adaptive, or disabled
Tools and output
Agent and computer-use workflows through MiniMax products and API

Limitations to plan around

  • Official pricing doubles above 512K input tokens.
  • Self-hosting requires substantial GPU capacity despite sparse activation.

Continue comparing

Related model decisions

See all reviewed models
Open-weight

Kimi K3

Long-horizon coding agents

Open-weight

DeepSeek V4

Cost-sensitive API workflows

Open source

Qwen3.6

Local coding assistance

Editorial note: This is a dated decision page, not a benchmark leaderboard or a substitute for legal, security, or infrastructure review. Model and pricing facts can change after October 23, 2026; use the linked primary sources before committing budget or customer data.Rechecked October 23, 2026: license, parameter count, context window, and API price. Modality, runtime, and reasoning notes date from the August 11, 2026 audit.