Fast-Tier Promotion: Qwen3.6-35B-A3B-NVFP4

Publication redaction: The operator-specific network hostname was replaced with a public-safe placeholder. The observed workflow and result are unchanged.

Date: 2026-07-08

Decision

Promote nvidia/Qwen3.6-35B-A3B-NVFP4, served as qwen36-35b-a3b-nvfp4, to the production Fast tier on Primary Node.

This is the human-gated follow-up to the Fast-tier bakeoff finding: 2026-07-08-fast-tier-llm-bakeoff.md. The bakeoff task recommended promotion but did not auto-promote. This finding records the explicit promotion apply and validation evidence.

Promoted Runtime

Component	Value
Host	Primary Node
Production serve	fast / vllm-qwen36
Port	127.0.0.1:30003
Model	nvidia/Qwen3.6-35B-A3B-NVFP4
Served model name	qwen36-35b-a3b-nvfp4
Engine	vLLM nightly
Quantization	modelopt_fp4
Context	32768
Max sequences	2
Router preset	chat-fast -> fast-local

Mini remains model-free in the reference topology. OpenClaw Gateway and Anvil Voice Realtime/proxy run on Mini; STT/TTS/LLM model endpoints remain on Fakoli Dark or are reached through Mini proxy ports that forward to Dark.

Evidence

Bakeoff metrics from 2026-07-08-fast-tier-llm-bakeoff.md:

Candidate/config	Voice total / LLM stage	TTFT / E2E	Approx decode tok/s	Hard gates
nvidia/Qwen3.6-27B-NVFP4 control	1130.21 ms / 814.83 ms	6203.94 ms / 9041.91 ms	67.65	pass
nvidia/Qwen3.6-35B-A3B-NVFP4	377.52 ms / 165.40 ms	1489.36 ms / 2302.37 ms	236.16	pass

Direct Fast endpoint preflight after serve swap:

anvil-serving preflight --base-url http://127.0.0.1:30003/v1 --model qwen36-35b-a3b-nvfp4 --needle-ctx 32768 --tool-batch 5 --no-thinking

Result: all probes passed: smoke, structured JSON, 32768-token needle retrieval, and shared-prefix tool batch.

Router promotion command:

anvil-serving router promote --profile docs\findings\fast-tier-bakeoff-evidence\fast-tier-promotion-profile.json --config examples\primary-node\anvil-router.live.toml

Result: profile and config were written to the anvil-router-cfg volume and anvil-router restarted successfully with the promoted profile/config.

The promoted profile is an operational routing gate, not fresh per-model quality calibration for every work class. It keeps Fast eligible for low-latency chat/voice and bounded-edit paths, while denying explicit Fast pins for planning and review until those high-risk classes are recalibrated on qwen36-35b-a3b-nvfp4.

High-risk pin proof after the stricter profile was applied:

{
  "status": 200,
  "provider": "heavy-local",
  "model": "gpt-oss-120b",
  "tier": "local",
  "work_class": "planning",
  "reason": "pinned; quality gate: pin fast-local denied for planning; routed via gated pool"
}

Route proof from inside the authed anvil-router container:

{
  "status": 200,
  "provider": "fast-local",
  "model": "qwen36-35b-a3b-nvfp4",
  "tier": "local",
  "work_class": "chat-fast",
  "reason": "preset='chat-fast'; quality gate: on"
}

End-to-end router relay proof through POST /v1/chat/completions:

{
  "status": 200,
  "model": "chat-fast",
  "finish_reason": "stop",
  "content": "Fast tier"
}

Runtime status after promotion:

Serve	Container	Port	Health
Heavy	vllm-gptoss120	30002	200
Fast	vllm-qwen36	30003	200

OpenClaw gateway sync:

anvil-serving harness sync openclaw --config examples\primary-node\anvil-router.live.toml --base-url http://100.64.0.10:8000/v1 --voice --voice-realtime-url ws://127.0.0.1:8765/v1/realtime --gateway-host mini-host.example --restart

Result: synced six Anvil preset models into mini-host.example:~/.openclaw/openclaw.json, merged with existing config, took a backup, and restarted the OpenClaw gateway.

Local repo gates:

Gate	Result
python -m ruff check anvil_serving tests scripts	pass
python -m pytest tests/ -q	2086 passed, 2 skipped
docker compose -f examples\primary-node\docker-compose.yml config --quiet	pass
anvil-serving serves --manifest examples\primary-node\serves.toml --dry-run up fast	pass
MCP openclaw_sync preview with voice=true	voice_model=chat-fast, voice_consult_model=anvil/chat-fast

Apply Notes

The first live router promotion attempt rolled back correctly because the Windows checkout wrote CRLF line endings into /etc/anvil/config.toml; the deployed router image rejected literal \r characters in TOML. The promotion path now normalizes text written into the Linux config volume to LF and writes stdin with UTF-8 encoding. Regression coverage lives in tests/test_router_manage.py.

The promoted repo state also updates OpenClaw/voice manifests to call the chat-fast preset and expect route proof for fast-local / qwen36-35b-a3b-nvfp4 / local.

Adversarial Review Follow-up

Two read-only review passes were run after the initial apply. The runtime/config review found that MCP openclaw_sync still defaulted Talk to the tier pin fast-local; the MCP wrapper now preserves explicit voice_model values but otherwise lets the harness resolve the preset default (chat-fast, then chat). The same review found that the promotion profile still allowed pinned Fast planning; the profile now denies Fast planning and review until those classes are recalibrated on the promoted model.

The docs/evidence review found that the profile could be misread as fresh per-model quality calibration; this report now labels it as an operational routing gate. It also found a stale 64K rerun command and an old setup story that read like current guidance; both were corrected.