This content presents benchmark results for the MiniMax-M2.7 LLM, utilizing NVFP4 quantization on a dual NVIDIA RTX PRO 6000 Blackwell GPU setup. It details decode throughput at various concurrency levels and prefill performance across different context sizes.
An investigation into MiniMax-M2.7 GGUF revealed that perplexity NaNs affect 21-38% of GGUFs on Hugging Face. The issue was traced to overflowing in llama.cpp, specifically in `blk.61.ffn_down_exps` for Q5_K and Q4_K quantizations, and the team has fixed theirs.
The MiniMax M2.7 model is criticized for having open weights but a restrictive license that prohibits commercial and military use. This approach frustrates the AI community, which sees a worrying trend of 'open weights, closed license'.
MiniMax M2.7 is a competitive 230-billion-parameter Mixture-of-Experts AI model, released in March 2026. Designed for "agentic" workflows, it delivers capabilities approaching proprietary competitors. The model maintains significantly lower operational expenses for organizations implementing agent-based systems.
Open models like GLM-5 and MiniMax M2.7 have crossed a threshold, matching closed frontier models on core agent tasks. They offer comparable performance in file operations, tool use, and instruction following, but at a fraction of the cost and latency.
