The Assembled Worker

Test the model you can actually run.

HMIC measures local models as deployed systems: source, quant, hardware, runtime, speed, task accuracy, control, and the operator guidance required to make them useful.

Published reports

Qwen3.6-35B versus Qwen3.8-27B benchmark

Benchmark 01 / RTX 5090

Qwen3.8-27B versus Qwen3.6-35B-A3B

35B-A3B ran 3.82 times faster in the measured deployments. 27B remained the safer untouched starting point. One compact operating card moved 35B from 12/24 to 16/24 on an unseen holdout.

48 timing calls72 holdout callsartifact hashesclaim boundary

Read the report Watch the analysis

Now running

Q4 versus Q6

One pinned model source
RTX 5090, one RTX 3090, and two RTX 3090s
Identical context, runtime settings, prompts, and scoring
Speed, TTFT, memory, semantics, and exact format

Bring a real failure.

Send the exact model, source, quant, hardware, context, runtime, prompt, expected behavior, and observed failure. Useful requests become candidate benchmark cells.

Request a test

Follow the series

The website carries tables, methods, limits, and durable corrections. The YouTube channel carries the narrated analysis and the finished visual evidence.

Open The Assembled Worker