The End of the Cloud

Connected 2 Asus Ascent GX10 units to expand a local inference setup, and directly tested several models for agentic coding.

• Initially felt that 1 unit with 128GB wasn't enough, so added a second GX10.
• Models tested were Qwen 3.5 122B-A10B, Qwen3-Coder-Next, M2.5-REAP, Qwen 3.5 397B-A17B, MiniMax M2.5 AWQ, and MiniMax M2.7 AWQ.
• Qwen 3.5 397B-A17B wasn't bad, but fell short as a reliable agentic teammate, and tended to declare results finished too quickly.
• MiniMax M2.5 AWQ is slow and has no vision, but was a very stable workhorse for agentic tasks, and MiniMax M2.7 was rated as an even better fit.
• In particular, when hooked up to a verification loop with tests or playwright-cli, it produced good results in planning, issue understanding, feature development, and bug fixing.
• However, one shouldn't expect the thoroughness of GPT-5.4 level or the strong watchdog tendencies of Opus 4.6, and that gap needs to be accounted for.

In conclusion, the author judged that the MiniMax M2.7 AWQ + 2 GX10 combo is enough to build a sufficiently practical agentic coding environment locally, and says they're now relying less on cloud providers.