local-llm
- MTP acceptance is a side effect, not a target
A friend said MTP acceptance near 0.8 meant an idle GPU and I should draft more. I swept draft length on a 27B. Acceptance fell; speed peaked early.
- Three bugs between K2 and a working Spring app
Two models, 0 lie cells each, same sampler: Swift-1.5 shipped both Java tasks, K2-Horizon neither. Its Spring app never booted; three patches fixed it.
- The model said it saw no output. It was right.
North-Mini scored 29 in every Claude Code cell and quit after about a minute. It named every tool call 1, and Claude Code passed it one result.
- Claude Code read my home folder as a project
The bench's Claude Code cell moved its config dir and excluded the repo's CLAUDE.md. My global ~/.claude still loaded, labelled as project instructions.
- The model card said think harder. The audits got worse
Qwen3.8's card defaults to xhigh and warns lower effort can mean more failures. xhigh passed 1 of 24 audits and 24 of 36 coding runs, against 4 and 27.
- It fits in 16 GB. That was the easy question
Qwen3.8-27B with its KV cache in system RAM fits a 16 GB card. Generation drops from 42 to 15 tokens/s at once, and to 1.5 at 200k context.
- The hard tier sorted by a rollback, not by speed
Five fast MoEs passed 1 of 45 hard-tier runs and skipped the same rollback in 33 of 45 ledger runs. A fast 2-bit model passed 7 of 9.
- The hidden test asked for more than the prompt
39 agent runs on a Java ledger task were marked down by one hidden check. The prompt never asked for it, and the project's README steered them into failing it.
- Empty diff, finished code: 124,000 characters of thinking
Two K2-Horizon agent runs scored 29 with an empty diff. The code was in the reasoning: the Go cell's, pasted into the repo, passes 27 of 27 hidden tests.
- The harness decided how long the model thought
One 27B model, one llama.cpp server: 250 s per agent run in OpenCode, 116 s in Pi. Under OpenCode the model reasoned 3.6 times as much.