Publication summary: ThinkingCap Qwen3.8 27B AWQ 32K BF16 qualification¶
Canonical facts¶
- Model identity:
bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ@f8fe157f207a13f977bc3d620ce10a3e9ba5ab11. - Runtime identity: vLLM 0.29.0; Triton attention, MTP3, BF16 KV.
- Recipe: 32K BF16 snapshot.
- Measurement path: direct managed C1 endpoint.
- Capacity: median 455.4421 ms TTFT and 158.355179 tok/s decode; 61.14585 tok/s aggregate, n=100 strict. Warm process, unique prompts, prefix cache off, nominal 4K input, 45 output tokens, thinking off. Whole-recipe result; no isolated MTP speedup claim.
- Qualification: 27 functional, 8 thinking, 18/18 vision, 30/30 limited MMLU, context 60/60, agentic 2/2.
- Boundary: two images about 1 MP; zero video; no C2, SWE, soak, maximum-context, route, or promotion claim.
- Evidence: finding, context, agentic.
X / short post¶
Local RTX 5090: ThinkingCap Qwen3.8 27B AWQ qualified a bounded 32K/C1 Triton MTP3 BF16 direct lane: 158.355 tok/s median decode, 60/60 context, 2/2 agentic. No promotion: docs/findings/2026-09-23-thinkingcap-5090.md
Reddit¶
One RTX 5090, direct 32K/C1 Triton + MTP3 BF16 KV: 100 strict short-output requests measured 455.4421 ms median TTFT, 158.355179 tok/s median decode, and 61.14585 tok/s aggregate throughput. Standard functional, vision, limited quality, 60/60 context, and 2/2 agentic gates passed. No route or promotion changed.
Screenshot alt text¶
A 32K/C1 Triton + MTP3 BF16 KV local result on one RTX 5090: 158.355179 tok/s median decode, 60/60 context, and 2/2 agentic. The scope excludes C2, video, SWE, soak, and promotion.
Claim ledger¶
| Public claim | Conditions | Evidence |
|---|---|---|
| 32K capacity | n=100 strict, nominal 4K/C1 | native |
| Context | 60/60, retained actual-token ranges | context |
| Agentic | 2/2, six turns, four tools | agentic |
| Final health | pinned 32K managed state, direct retention authorized | ending state, HTTP, managed status |
| No promotion | no route/client alias/deployed-topology change | finding |