Skip to content

ThinkingCap Qwen3.6-27B FP8 Heavy promotion

Point-in-time record, 2026-07-12. ThinkingCap Qwen3.6-27B FP8 was promoted from resident Heavy challenger to the routed heavy-local default on Fakoli Dark after a repaired, model-aware preflight. GPT-OSS-120B was stopped and retained as the complete serve-and-router rollback state. Fast and Voice were not changed.

Deployed recipe

Field Promoted value
Served model thinkingcap-qwen36-27b-fp8
Checkpoint bottlecapai/ThinkingCap-Qwen3.6-27B-FP8 revision e48255afd77b403446332be0f595868337b36591
Host / accelerator Primary Node / one RTX PRO 6000 Blackwell 96 GB
Engine / quantization pinned vLLM image digest sha256:907377dd...5319ff3e; FP8 weights and FP8 KV
Speculative decoding Qwen3 MTP, 3 speculative tokens
Context / admission serve 262,144 tokens and five sequences; router advertises a conservative 131,072-token Heavy window
Heavy default thinking enabled through extra_body_defaults
Completion allocation 256 visible-answer tokens plus 4,096 reasoning-headroom tokens, sent as a 4,352-token cap

The selected 4K operating point comes from the repeated protocol-v2 result: ThinkingCap retained 15/15 ARC attempts and reached 9/10 stable MMLU-Pro items with 27/30 passing attempts. The broader comparison, five-session capacity result, recipe caveats, and source lineage remain in the Qwen3.6 protocol-v2 finding.

Promotion gates

The functional gate explicitly disabled thinking and used a 256-token cap. It passed short coding, structured JSON, a 131,072-token needle, and 20/20 shared-prefix tool calls. Every response ended with stop or tool_calls, and no parsed reasoning characters or reasoning-token usage were observed. See functional-preflight.json.

The separate quality smoke explicitly enabled thinking with 256 visible tokens plus 4,096 reasoning headroom. Coding and JSON both ended with stop, produced visible answers, and exposed 1,246 and 639 parsed reasoning characters. See thinking-preflight-4k.json.

Both the forward and rollback profile/config pairs passed the rebuilt deployed router image's loader before the model swap. After profile/config replacement, the router reloaded successfully and the actual Tailnet health endpoint http://100.64.0.10:8000/healthz returned HTTP 200. The direct Heavy health endpoint at http://127.0.0.1:30002/health also returned HTTP 200. The complete guarded operation finished in 307.1 seconds.

Safety and caveats

Preflight now records full visible output, finish_reason, parsed reasoning field/length/excerpt, reasoning-token usage when exposed, and indexed tool-call validation. It rejects disallowed finish reasons and can require or forbid effective reasoning evidence. The promotion plan requires both gates before router mutation and automatically restores GPT-OSS on failure; its cold-start rollback timeout is 1,200 seconds because the earlier 600-second value was not safe on this host.

This promotion does not establish five simultaneous 262K sessions or a one-million-token operating window. Five-session evidence used independent 8K contexts, and the routed Heavy context remains 131K pending retained long-window validation. GPT-OSS remains the rollback model, not the Heavy default.

Raw artifact hashes are recorded in SHA256SUMS.