ThinkingCap 27B 128K secondary promotion — 2026-09-24¶
The user explicitly approved promotion of the qualified RTX 5090 ThinkingCap profile. The managed router install passed with the expected served identity, and the existing llm.secondary alias now advertises 131,072 total tokens, 10,240 maximum output tokens, one active request, and up to two images. Fleet-client and observability completion are tracked separately below.
Local outcome: user-authorized router promotion; independent review accepted the exact qualified recipe.
Setup: one RTX 5090 (32,607 MiB), Windows/Docker Desktop/WSL2 model host, a separate native Linux router, vLLM 0.29.0, NVFP4A4 AWQ, BF16 KV, MTP3 and generic UVA3. Warm routed online requests; this is a deployment check, not a new performance benchmark.
Checks: seven C1 functional gates and two enabled-reasoning gates passed at the initial output setting. At the final 10,240-token setting, five text/tool gates passed; two failed multimodal calls passed on direct diagnosis and routed retry without mutation. The measured startup KV allocation is 138,519 tokens, distinct from the 131,072-token context contract.
Limit: a C20 burst produced six successful tool responses and fourteen 30-second admission timeouts. The qualified contract remains C1. Near-limit text and small images were qualified separately.
Identity and scope¶
The model is bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ at revision f8fe157f207a13f977bc3d620ce10a3e9ba5ab11, served as thinkingcap-awq-5090. The vLLM image digest is sha256:082ca6f035279109041ffd3fe0695cb568b29bc580b35c4f297a66a08b216c1b, engine revision 98dff2a81d747d1dba01a47f939f48c3526d4206, and exact recipe semantic digest c81bf6a3c79e6062c99258243d50a1c6feaa0886c8ff6724f79b2cca8f46f026.
The 128K qualification measurements and decisions and managed recipe reconstruction remain unchanged. Before publication, the public qualification bundles were resealed after container-ID redaction; their original seals are archived privately. The 32K and 128K ledgers record the exact transforms. This packet changes the dated decision from challenger to current; it does not rewrite prior no-promotion evidence. The earlier 32K recipe remains the faster qualified alternative. Only the RTX 5090 was measured; the router's separate model hardware was not benchmarked.
Router installation and verification¶
The exact installed router configuration was captured before mutation. The secondary-only change preserves primary, audio and alias mappings. Managed preview and confirmed installation applied the reviewed bytes; both tiers returned ready with verified identities. The private edge now forwards the existing model path to the retained qualified container. No model weights or launch settings changed during promotion.
The authenticated capacity, capabilities, and fingerprints distinguish configured contract, observed identity and live capacity. Independent review accepted the exact pins and secondary-only scope. The final managed install additionally declares the supported reasoning efforts: low, medium, and xhigh.
The initial 4,096-token output cap was corrected to 10,240: the sealed 30/30 diagnostic used that allowance, and three successful attempts exceeded 4,096 completion tokens. The budget audit and independent correction review retain the reason. Context is the total input-plus-output budget; the earlier near-limit context matrix used a 4,096-token reserve and does not prove 10,240 output tokens after a 125K-token prompt.
The explicit C1 routed preflight passes smoke, JSON, one shared-prefix tool request, streaming tools, tool-result continuation, image and OCR. The enabled-reasoning check passes both requested gates with reasoning-channel evidence.
Fleet and observability convergence¶
Router promotion is complete. Client catalogs and scheduled reconciliation are tracked in the convergence record. A real Pi request exposed a reasoning compatibility gap described below. After an explicit per-model high to xhigh mapping, both an isolated Pi check and the normal installed Pi environment passed a shell-tool/result conversation. The independent event checks validate one successful tool execution and exact final fixture contents. The normal-environment router trace independently records two successful requests to the secondary tier, ending in tool_calls and stop, with the final configuration identity and 10,240-token allowance.
Discovered managed catalogs across the four operator hosts now carry the final limits. Pi and OpenClaw secondary display names identify ThinkingCap Qwen3.8 27B AWQ while retaining llm.secondary as the request alias. Companion scheduled reconciliation is approved for the final router contract and idempotent. Existing OpenClaw/Hermes acceptance is retained separately from final catalog refreshes. The Windows model host also passed native Pi high-effort read-tool/result acceptance. The durable catalog mapper passed 34 focused tests and independent review. Existing interactive Pi sessions can use /reload to refresh the model registry without resetting the conversation.
Open WebUI's persistent provider already references the router. A read using its configured provider credential from inside the application container returned llm.secondary with 131,072 context and 10,240 output tokens. No global generation override or cached model list was configured. Browser chat acceptance remains untested because interactive sign-in was not completed.
Prometheus now scrapes authenticated router metrics through the existing private HTTPS edge. Native queries returned context 131,072, loaded 1, image limit 2, and target up 1. Grafana's persisted dashboard resource contains the ThinkingCap 27B row and the three llm.secondary panels. This verifies router contract/readiness monitoring; the separate legacy observer does not yet provide the new remote engine's detailed throughput metrics.
Failures and limits¶
The first routed preflight inherited a default 20-way tool burst. The router's authoritative logs identify admission-deadline exhaustion for fourteen requests at 30 seconds. The failed native artifact is retained. An explicit one-request run passed. No C20 capacity or broad concurrency claim follows.
At the final output cap, image and OCR requests returned HTTP 500 / RelayBackendError; the engine remained healthy. Both direct diagnostic calls and routed retries passed without a model, route, or configuration change. The failed run remains evidence of an unresolved intermittent relay incident, not a fully clean run. Independent review accepted retaining the exact deployment with this incident recorded; missing upstream failure detail is an open product issue.
A real Pi request reached the intended secondary engine but returned HTTP 400. The metadata-only request trace and engine error identify reasoning_effort=high, which the pinned template rejects. Its supported values are low, medium, and xhigh. This client compatibility failure was corrected with an explicit per-model map, preserving the user's preferred high selection. Two initial verification processes waited on inherited SSH standard input and timed out before submitting model requests; the corrected launcher closes standard input. Those launcher failures remain separate from model acceptance.
A whole-home configuration export refused an unrelated secret-file symlink. A narrow installed-router export surface was added and independently reviewed; it validates exact read-only mounted bytes using the deployed parser and preserves the broader export guard. Its final focused regression set passed 155 tests with one skip. Separate service ownership and scheduled-contract gaps are tracked in the friction log.
No new broad-quality, exact-limit/one-over, mixed near-limit vision, video, SWE or soak qualification is claimed. Previous direct speed measurements remain workload-specific. Private host names, addresses and paths are replaced by synthetic identifiers in this packet; exact private rollback state remains with the operator.