Qwen3.6-35B-A3B-Q6_K Hermes Benchmark
The benchmark was started cleanly against the Q6_K local model and made meaningful progress. It was stopped after Hermes crossed the agreed maximum of three actual errors, including repeated fetch failures and a failed attempt to write the final artifact.
Run Summary
qwen/qwen3.6-35b-a3bqwen3.6-35b-a3b20260527_050343_bfd17dOutcome
This was not a speed failure like the prior low-quant MTP run. The model processed prompts and generated tool plans quickly enough to be usable. The failure condition was operational reliability: Hermes accumulated more than three real tool errors before completing the benchmark and then failed while attempting to write hermes_local_model_agent_benchmark_submission.html.
Timing
| Event | Observed time | Detail |
|---|---|---|
| Model confirmed | 2026-05-27 05:03 +0600 | lms ps and lms status showed qwen/qwen3.6-35b-a3b loaded in LM Studio with a 262144-token context and IDLE before the clean run started. |
| Hermes started | 2026-05-27 05:03 +0600 | New Hermes session 20260527_050343_bfd17d opened on qwen3.6-35b-a3b. |
| Reasoning set | 2026-05-27 05:03 +0600 | Hermes confirmed: reasoning effort set to high. |
| Benchmark prompt submitted | 2026-05-27 05:03:54 +0600 | Hermes was instructed to read /Users/armanshawon/Documents/Benchmark/test-prompt.md and create the output inside /Users/armanshawon/Documents/Benchmark. |
| Initial prompt processing | 05:03:54 to about 05:04:32 | LM Studio logs showed prompt processing progressing from 0.0% to 95.7% in roughly 38 seconds. |
| Second model call processing | 05:04:58 to 05:05:02 | LM Studio logs showed a later 6-message request processing from 0.0% to 100.0% in about 4 seconds. |
| Run stopped | 2026-05-27 05:10 +0600 | Hermes was interrupted and exited after the error threshold was exceeded. Exit summary reported Duration: 6m 10s and Messages: 51 (1 user, 50 tool calls). |
| Post-run model state | 2026-05-27 05:10 +0600 | LM Studio still had qwen/qwen3.6-35b-a3b loaded and IDLE. |
Tool Call Detail
| Tool / action | Observed result | Benchmark relevance |
|---|---|---|
read benchmark prompt |
Successful. Hermes read test-prompt.md in about 1.1s, then read it again in about 0.8s. |
Satisfies prompt ingestion and shows the run started correctly. |
execute_code / Python |
Successful. Hermes ran Python setup/import work in about 3.2s. | Indicates the model began the required calculation workflow. |
| Web searches | Several searches succeeded, including local LLM candidates, coding-agent models, and affiliate program validation queries. Observed durations ranged from about 1.7s to 3.8s. | Matches the live research portion of the benchmark. |
| Browser navigation and snapshots | Successful navigation and scrolling were observed on pages such as dev.to, xda-developers.com, and netpeaksoftware.com. | Shows the model could perform multi-step web inspection, not only search. |
| Tool name auto-repair | Hermes auto-repaired browser_extract to web_extract twice. |
Non-fatal compatibility mapping. Not counted as a benchmark error by itself. |
| Fetch failures | Four actual fetch-level errors were observed: huggingface.co +2, two separate pages fetch errors, and esimradar.com +1. |
The fourth fetch-level error exceeded the user's maximum tolerance of three errors. |
write final HTML |
Hermes attempted to write /Users/armanshawon/Documents/Benchmark/hermes_local_model_agent_benchmark_submission.html, but the write call returned [error]. |
This prevented a valid benchmark artifact from being produced. |
Assessment
Qwen3.6-35B-A3B-Q6_K looked materially more viable than the previous IQ2_M/MTP run on this machine. It reached tool use quickly, kept Hermes active, and handled a large context without the very slow first-response behavior. However, the run cannot be scored as a completed submission because the required HTML file was not written and the session exceeded the pre-set error budget.
Recommended classification: failed run due to tool reliability and final artifact write failure, not due to raw model speed.