NOUMENOpen app
NOUMEN

MiniMax · Open weights

MiniMax M3 Thinking

Added Jun 1, 2026Open weights

Compare models

Choose up to 3 models to compare.0 of 3 chosen

Loading models…

Use in the app

MiniMax M3 Thinking is the adaptive-thinking version of MiniMax's open-weights frontier model for coding, agent workflows, tool use, long-context tasks, and native multimodal understanding. MiniMax reports 59.0% on SWE-Bench Pro and 66.0% on Terminal Bench 2.1, with Sparse Attention designed to scale context to 1M.

  • Reasoning
  • Tools
  • Vision
  • Structured output
Input
Text, Images
Output
Text
Discovery tier
Open weights

Pricing

USD per 1M tokens
Input
$0.23
Output
$0.96
Cache read
$0.05

Reference prices as published by the model provider. API example

Specifications

Context window
512K
Max output
80K
Average output (7 days)
1.9K tokens
Longer than 90% of 465 models

Side by side

Compare models

Put MiniMax M3 Thinking next to up to three other models. The comparison stays in the link, so you can share it.

Compare models

Choose up to 3 models to compare.0 of 3 chosen

Loading models…

TryGemini Pro LatestGPT 6 LunaClaude Sonnet 4.6 ThinkingAll 3

Public evaluations

Benchmarks

Independent evaluation results, with how each score ranks among the models in the Noumen catalog.

Intelligence Index

29.2

Better than 70% of 326 scored models

3.8Catalog models by score55.1

Indexes

  • Coding Index

    58.6

    Composite of the coding evaluations

    Better than 58% of 168 scored models

Agentic work

  • AutomationBench-AA

    21.3%

    Workflow automation, penalized for guardrail violations

    Better than 43% of 173 scored models

  • AutomationBench-AA tasks completed

    4.4%

    Workflows finished end to end without a guardrail violation

    Better than 45% of 173 scored models

  • Harvey LAB-AAv1.1

    0.0%

    Legal tasks passed in full, with no material hallucination

    Better than 0% of 29 scored models

  • AA-Briefcasev1.1

    1,089 Elo

    Agentic knowledge work, rated head to head

    Better than 49% of 166 scored models

  • GDPval-AAv2.1

    1,243 Elo

    Economically valuable occupational tasks, rated head to head

    Better than 54% of 187 scored models

Document reasoning

  • GDP.pdf

    9.8%

    Professional reasoning over PDF documents, all-pass rate

    Better than 36% of 171 scored models

  • AA-LCRv1.1

    83.0%

    Reasoning across long documents

    Better than 90% of 309 scored models

  • MLCR-AA

    4.4%

    Long-context reasoning over medical records

    Better than 10% of 121 scored models

Reasoning and knowledge

  • Humanity’s Last Exam

    39.0%

    Expert-level questions across academic fields

    Better than 73% of 323 scored models

  • GPQA Diamond

    92.9%

    Graduate-level science questions

    Better than 91% of 291 scored models

  • IFBench

    82.9%

    Following precise output instructions

    Better than 99% of 215 scored models

Coding

  • Terminal-Benchv4.0

    2.0%

    Practical software engineering in a terminal

    Better than 45% of 171 scored models

  • SciCode

    47.1%

    Scientific computing in Python

    Better than 39% of 174 scored models

Legacy evaluations

  • Terminal-Bench Hard

    42.4%

    Earlier terminal task suite

    Better than 79% of 212 scored models

  • τ²-Bench Telecom

    88.9%

    Tool use in a customer-support simulation

    Better than 72% of 214 scored models

  • AA-LCR (unversioned)

    83.0%

    Long-context reasoning, earlier grading

    Better than 90% of 309 scored models

Scores from Artificial Analysis, updated Oct 26, 2026. Artificial Analysis

Live routing data

Performance and reliability

Recent averages across requests routed for this model. Your own speed depends on prompt length, reasoning effort and load.

Throughput
226.1 tok/s
Faster than 87% of 247 models
Time to first token
1.7 s
Quicker than 85% of 247 models
Prompt cache hits
56.4%
About $0.13 per 1M input tokens with caching
Request success (24h)
99.9%
Reliable

Cache hits are the share of input tokens recently served from a prompt cache across all requests. They are not a forecast of your own hit rate.

Hourly request success · last 72 hoursLowest hour 87.5%
  • Healthy (98% or more)
  • Degraded (90–98%)
  • Down (under 90%)
  • No requests

Developers

Use it through the API

Noumen’s API is OpenAI-compatible: any Chat Completions client works with a base URL and a key.

Read the quickstart
curl https://www.noumen.xyz/api/v1/chat/completions \
  -H "Authorization: Bearer $NOUMEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/minimax-m3:thinking",
    "messages": [{ "role": "user", "content": "Explain this stack trace." }]
  }'
import os
from openai import OpenAI

client = OpenAI(base_url="https://www.noumen.xyz/api/v1", api_key=os.environ["NOUMEN_API_KEY"])
reply = client.chat.completions.create(
    model="minimax/minimax-m3:thinking",
    messages=[{"role": "user", "content": "Explain this stack trace."}],
)
print(reply.choices[0].message.content)

MiniMax

Other MiniMax models, starting with the same family.