---
license: other
pretty_name: Qwen3.6 MoE Coding Expert Saliency Analysis
tags:
- qwen
- moe
- llama-cpp
- saliency
- terminal-bench
- scicode
- apple-silicon
---

# Qwen3.6 MoE Coding Expert Saliency Analysis

This repository contains the combined coding-task expert-saliency report for local Qwen3.6-35B-A3B MoE tracing work.

Primary artifact:

- [`combined_coding_expert_analysis.html`](combined_coding_expert_analysis.html): combined expert activation, router-weight, contribution-proxy, correlation, and residency-candidate analysis across the SciCode MVP trace set and the full Terminal-Bench Hard real-agent trace run.

Companion artifacts:

- [`combined_coding_expert_residency_candidates.json`](combined_coding_expert_residency_candidates.json): machine-readable observational residency candidates.
- [`reports/scicode_expert_saliency_analysis.html`](reports/scicode_expert_saliency_analysis.html): source SciCode report.
- [`reports/terminal_bench_hard_agent_saliency_analysis.html`](reports/terminal_bench_hard_agent_saliency_analysis.html): source Terminal-Bench Hard agent report.
- [`summaries/scicode_saliency_summary.json`](summaries/scicode_saliency_summary.json): SciCode summary metrics.
- [`summaries/terminal_bench_hard_agent_saliency_summary.json`](summaries/terminal_bench_hard_agent_saliency_summary.json): Terminal-Bench Hard agent summary metrics.
- [`recommendations/atx_expert_residency_policy.json`](recommendations/atx_expert_residency_policy.json): local ATX residency policy artifact, when available.
- [`upload_manifest.json`](upload_manifest.json): file hashes, source paths, and upload metadata.


## Portable ATX Residency Policies

The `policies/` directory contains four portable `--moe-residency-policy` JSON files generated from the combined coding saliency data:

| File | Selected layer-expert cells | Combined importance share |
|---|---:|---:|
| [`policies/combined_top_05pct_layer_experts.atx.json`](policies/combined_top_05pct_layer_experts.atx.json) | 512 | 66.1% |
| [`policies/combined_top_10pct_layer_experts.atx.json`](policies/combined_top_10pct_layer_experts.atx.json) | 1,024 | 80.0% |
| [`policies/combined_top_15pct_layer_experts.atx.json`](policies/combined_top_15pct_layer_experts.atx.json) | 1,536 | 87.1% |
| [`policies/combined_top_20pct_layer_experts.atx.json`](policies/combined_top_20pct_layer_experts.atx.json) | 2,048 | 91.2% |

Example usage with the ATX fork:

```bash
./build-atx-metal/bin/llama-server \
  -m /path/to/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf \
  --moe-residency-policy policies/combined_top_10pct_layer_experts.atx.json \
  --moe-residency-stats stats.json \
  --ctx-size 4096 \
  --parallel 1
```

These policies rank layer-expert cells by mean normalized importance across SciCode MVP and Terminal-Bench Hard Agent traces. They recommend which cells to duplicate/cache hot in VRAM; they are not yet throughput-validated as globally optimal.

## Important Caveats

These results are observational routing/contribution-proxy measurements. They are useful for proposing which experts or layer-expert cells should remain VRAM-resident, but they are not a causal proof that those policies preserve quality or maximize throughput. Final residency recommendations should be validated with forced-residency/offload runs on the same prompts or agent tasks.

This upload intentionally excludes raw trace files and large parquet intermediates. It is a shareable analysis bundle, not the complete trace corpus.

The reports are not official Artificial Analysis benchmark reproductions. They use public/reconstructable tasks and local instrumentation to study MoE routing behavior.