NVIDIA NIM MiniMax M3
MiniMax M3 1M MSA Sparse Attention model on NVIDIA NIM — 59% SWE-bench Pro for autonomous coding agents.
Updated 8/29/20261 code examples
About the Model
Why Choose MiniMax M3?
MiniMax M3 is MiniMax's flagship full-stack code and reasoning model for long-range autonomous agents and massive refactors. It scored 59.0% on SWE-bench Pro in 2026, just behind Kimi K3 and ahead of DeepSeek V4.
Key Strengths
- 1M MSA sparse attention: ingests 100k-line repos or massive docs in one shot with low latency.
- Top end-to-end coding and agent planning: native tool calling and self-correction, 66.0% on Terminal-Bench 2.1.
- Native multimodal: analyzes images and UI recordings to spot architecture and UI bugs and generate fixes.
- 14 days 100% free on GMI Cloud MiniMax Week, no credit card required.
How to Access for Free (via NVIDIA NIM)
Why Choose MiniMax M3 on NVIDIA NIM?
MiniMax M3 on NVIDIA NIM provides GPU-accelerated inference for long-horizon autonomous software engineering and 100k-line repo refactoring with 1M context.
Try it in your browser
Pick a model, paste your own free key, and run. Your key is sent once to call the provider and never stored on our servers.
Pick a model, paste your own free key, and run. Your key is sent once to call the provider and never stored on our servers.
Code Examples
curl
curl https://integrate.api.nvidia.com/v1/chat/completions \
-H "Authorization: Bearer $NVIDIA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimaxai/minimax-m3",
"messages": [{"role": "user", "content": "Review this entire codebase for race conditions."}]
}'
Community Feedback
Comments are tied to your GitHub account — sign in to join the discussion.