Home / Models / MiniMax M3

MiniMax · MiniMax Community

MiniMax M3 API

Economical 1M-context multimodality with sparse attention.

Get API keyCompare models

Context: 1M tokens · Released: June 2026

# MiniMax M3 via kryyt (OpenAI SDK)
from openai import OpenAI
client = OpenAI(base_url="https://api.kryyt.com/v1", api_key="$KRYYT_API_KEY")
r = client.chat.completions.create(
    model="minimax-m3",
    messages=[{"role":"user","content":"Explain MoE routing in one paragraph."}],
)
print(r.choices[0].message.content)

Overview

MiniMax M3 is a 428B-parameter mixture-of-experts model (~23B active) using MiniMax Sparse Attention to make a 1M-token context economical to serve. It takes text, image, and video natively and is strong on agentic coding — an efficient default for long agent loops.

Best for:

  • Economical long-context agents
  • Native image + video input
  • Coding and tool use at scale
MiniMax M3Example Kryyt AI workspace with chat, project context, and linked files

MiniMax M3 benchmark performance

MiniMax M3 targets efficiency: strong web-research and tool-use results for a model with ~23B active parameters. MiniMax has not published GPQA or HLE scores for M3.

MiniMax M3 evaluation results
BenchmarkModelScore (%)
BrowseCompAutonomous web researchMiniMax M3
83.5
MCP AtlasTool use through MCP serversMiniMax M3
74.2
Terminal-Bench 2.1Agentic coding in real terminal environmentsMiniMax M3
66.0
SWE-bench ProResolving issues in professional codebasesMiniMax M3
59.0

Sources: MiniMax M3 launch coverage (VentureBeat) (opens in new tab).

MiniMax M3 vs MiniMax M2.7

Generation-over-generation: MiniMax M2.7 → MiniMax M3
BenchmarkScores (%)
SWE-bench ProResolving issues in professional codebases
MiniMax M2.756.2
MiniMax M359.0

Sources: MiniMax-M2 series technical report (opens in new tab); MiniMax M3 launch coverage (VentureBeat) (opens in new tab).

Independent check: Terminal-Bench 2.1 for MiniMax M3 is 66.0 vendor-reported vs 53.6 measured by Vals AI on its own harness (source (opens in new tab)).

Scores are vendor-reported unless marked independent, as of September 2026. Labs use different agent harnesses, reasoning settings, and benchmark versions, so treat cross-lab gaps of a few points as roughly equal and test on your own workload. See all models on the benchmarks page.

MiniMax M3 models & pricing

ModelAPI IDContextPricingLicenseStatus
MiniMax M3minimax-m31MSee account pricingMiniMax CommunitySoon
MiniMax M2.5minimax-m2.5204KSee account pricingMiniMax CommunitySoon

minimax-m3

Model
MiniMax M3
Context
1M
Pricing
See account pricing
License
MiniMax Community
Status
Soon

minimax-m2.5

Model
MiniMax M2.5
Context
204K
Pricing
See account pricing
License
MiniMax Community
Status
Soon

Usage is charged at Kryyt's published rates for the selected model and features. Review current pricing and usage in your account.

MiniMax M3 FAQ

What is MiniMax M3?

MiniMax M3 is a 428B-parameter mixture-of-experts model (~23B active) using MiniMax Sparse Attention to make a 1M-token context economical to serve. It takes text, image, and video natively and is strong on agentic coding — an efficient default for long agent loops.

How do I call MiniMax M3 on kryyt?

Send a POST request to https://api.kryyt.com/v1/chat/completions with your kryyt key and set "model" to one of: minimax-m3, minimax-m2.5.

What license is MiniMax M3 under?

MiniMax M3 is available under the MiniMax Community license via its provider, MiniMax.

What context window does MiniMax M3 support?

MiniMax M3 supports 1M tokens depending on the specific model variant.

Related models

DeepSeek

DeepSeek V4.1

Frontier open reasoning and agentic coding at commodity prices.

Z.ai (Zhipu)

GLM 5.3

Agentic engineering — long-horizon coding, terminal tasks, hybrid reasoning.

Moonshot AI

Kimi K3

The largest open-weights model — native vision and multi-agent Swarm.

Ship with one endpoint

Get a key, change your base URL, and call every model on this page. No new SDKs.

Get your API key