Falls Behind M2.5

In a total of 445 trials on Terminal-Bench 2.0, running MiniMax-M2.7 (Q8_0, unsloth GGUF) locally on a Mac Studio M3 Ultra 512GB produced an average performance of 41.3%.

• 184 solved / 250 failed
• Of 198 errors, 187 were AgentTimeoutError
• Generation speed was 10-17 tok/s

On the same hardware and configuration, the previous version M2.5 recorded 42.7%, meaning M2.7 actually scored lower. Timeouts also increased from 166 to 187, interpreted as tasks failing to complete within the time limit due to slower generation speed.

The execution environment consisted of llama.cpp build 8680, Metal GPU offload, and a claude-proxy placed between Claude Code and llama-server. llama-server received the Anthropic Messages API directly, running the local model as an agent.

The key conclusion is that M2.7 did not show a clear improvement over M2.5 for agentic coding on Apple Silicon. Combined with licensing confusion at the time of release, the incentive for further investment was weak.