Close
A standalone PowerShell module provides the fastest route to local installation.
Refer to the action plan below to initialize the model.
All large files and heavy weights are downloaded automatically by the script.
The deployment tool scans your environment and chooses the ideal parameters.
MiniMax-M2.7-NVFP4 is a highly optimized, 4-bit quantized variant of MiniMaxAI’s flagship 230-billion parameter sparse Mixture-of-Experts (MoE) foundation model, compressed via NVIDIA Model Optimizer using the cutting-edge NVFP4 (Nvidia Floating Point 4-bit) format. The architecture leverages a blockwise FP8 scaling scheme per 16 elements, dropping the previous Lightning Attention layers in favor of pure, hardware-optimized Grouped-Query Attention (GQA) with 48 query heads and 8 KV heads. This aggressive mathematical alignment allows the massive model to execute on a mere 10B active parameters per token, reducing VRAM demands dramatically down to 70 GB per GPU in Tensor Parallel setups. Tailored for self-evolving agent loops, multi-file code refactoring, and real-world system debugging, it delivers extreme processing throughput over an expansive 196,608-token context window while maintaining an exceptional 56.22% score on the SWE-Pro engineering benchmark.
| Specification | Detail |
|---|---|
| Total / Active Parameters | 230 Billion Total / 10 Billion Active per Token (Sparse MoE) |
| Quantization Layout | NVFP4 (4-bit Weights with Blockwise FP8 Scales via Nvidia Model Optimizer) |
| Context Window | 196,608 tokens (196k natively) |
| Hardware Baseline | Dual NVIDIA RTX PRO 6000 Blackwell (96GB GDDR7) or H100 Tensor Parallel |
| Attention Mechanism | Standard GQA Softmax (48 Query / 8 KV Heads) |
| Primary Execution Engines | vLLM Native Server, SGLang Backend with b12x |
| Core Benchmarks | SWE-Pro: 56.22% / Terminal Bench 2: 57.0% / VIBE-Pro: 55.6% |
