# ThinkingCap (Qwen3.8 27B) (slim)

> ThinkingCap-Qwen3.8-27B is BottleCap AI's fine-tune of Alibaba's Qwen3.8-27B that shortens its reasoning traces, a 27B dense vision-language model on Hugging Face under PolyForm Small Business 1.0.0, with a gated main repository.

- Full: https://www.anchorterminal.com/tools/thinkingcap-qwen3-8-27b.md (~10,650 tokens) · this version ~2,030 tokens · JSON https://www.anchorterminal.com/tools/thinkingcap-qwen3-8-27b.json · canonical https://www.anchorterminal.com/tools/thinkingcap-qwen3-8-27b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/thinkingcap-qwen3-8-27b#pricing · reviews https://www.anchorterminal.com/tools/thinkingcap-qwen3-8-27b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #50 of 100 open-weight models · #43 in Open-weight models · not agent-ready · confidence low**


Assessment: A Qwen3.8-27B with shorter reasoning, for agent work where thinking tokens cost time, and BottleCap's 4-bit GGUF runs in about 19 GB. The main limits are the licence, which needs a paid licence above 100 people or 1,000,000 USD revenue, a gated main repository, and no safety results or training-data description.

## Facts

- Kind: Open-weight model · vendor: BottleCap AI · category: Open-weight models · legal entity: BottleCap AI (named in the licence's copyright line) · provenance 47/100
- Packages: huggingface `bottlecapai/ThinkingCap-Qwen3.8-27B`, huggingface `bottlecapai/ThinkingCap-Qwen3.8-27B-FP8`, huggingface `bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF`, huggingface `bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4`, huggingface `bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ`, huggingface `bottlecapai/ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ`, huggingface `bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw`
- Auth: None · pricing: Free · x402: no · licence: PolyForm-Small-Business-1.0.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B, a gated repository, read on 21 October 2026 from its public card, the Hub's public metadata and file list, and BottleCap's ungated GGUF, NVFP4 and MLX repositories and blog post. The gate's terms weren't accepted, nothing was downloaded or run
- Model: A fine-tune of Qwen3.8-27B (base_model_relation finetune), Qwen3_5ForConditionalGeneration, 27,781,427,952 BF16 parameters per the Hub, the base's exact count. Text, images and video in, text out. The card says the aim was fewer thinking tokens at the same accuracy
- Files: 18 safetensors shards in BF16 (55.6 GB), chat_template.jinja, tokenizer files, LICENSE, the PolyForm and Apache 2.0 texts and a NOTICE. No pickle or Python files. The Hub hides the SHA-256 of LFS files from anyone who hasn't passed the gate
- Licence: PolyForm Small Business 1.0.0 plus BottleCap's grant for personal non-commercial use. Organisations with 100 or more people or 1,000,000 USD (2019) revenue need a commercial licence. NOTICE keeps Apache 2.0 for the upstream Qwen3.8-27B materials and names it as the base checkpoint
- Runtimes: The qwen3_5 architecture of the base runs in released llama.cpp, Ollama, vLLM, SGLang and Transformers. The card gives commands for Transformers, vLLM (qwen3_xml parser) and SGLang (qwen3_coder parser), and the GGUF card for llama.cpp, LM Studio and Ollama. No Ollama or LM Studio library build
- Context: 262,144 tokens, the base's native length, per the ungated NVFP4 build's config.json (max_position_embeddings 262144, rope_type default) and BottleCap's evaluation window
- Tools and reasoning: The same template as Qwen3.8-27B. Tools are declared in tools tags and calls come back as function and parameter blocks in tool_call tags. Thinking is on by default, reasoning_effort takes xhigh (default), medium or low, and the card shows the reasoning in a separate reasoning or reasoning_content field
- Hardware: The card evaluated on an NVIDIA H200 and lists build sizes (FP8 31 GB, GGUF 16 to 55 GB, NVFP4 21 GB, MLX 22.5 GB for a 32 GB Mac). The RCO GGUF card says 14.3 GB runs on an 18 GB GPU at 32K context
- Training data: Not described on the card or in the blog post

- Scores: Capability pending, Reach 75, Risk 45, Agent interface 70, Documentation and upkeep 53 · total over the 7 assessed categories
- Why:  · Reach, Memory 25 of 30, BottleCap's IQ4_XS GGUF is 15.5 GB, 18.6 GB with a fifth added, which fits 24 GB. · Risk, File safety 25, safetensors only and no code in the repository. · Agent interface, Chat template 20, the same Jinja template as Qwen3.8-27B, in the repository. · Documentation and upkeep, Model card 8 of 20, intended use, limitations and parameters are there, but architecture, context, languages and a cut-off aren't.
- Sources: 20, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/thinkingcap-qwen3-8-27b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/thinkingcap-qwen3-8-27b.svg` or a link to https://www.anchorterminal.com/tools/thinkingcap-qwen3-8-27b from a page on bottlecapai.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Check the licence first, an organisation with 100 or more people or 1,000,000 USD (2019) revenue needs a commercial licence from BottleCap
2. For files without the gate use `llama-server -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF:Q4_K_M --jinja`, about 17.4 GB
3. On vLLM use `--reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_xml`, and on SGLang `--tool-call-parser qwen3_coder`
4. Set `reasoning_effort` (xhigh, medium or low) in chat_template_kwargs, any other value makes the template raise an error
5. Keep the base's sampling (temperature 1.0, top_p 0.95, top_k 20), the GGUF card says greedy decoding can loop
6. The GGUFs carry an MTP head, and llama.cpp builds without MTP support for this architecture may refuse to load them with a missing blk.64 tensor

## Connect

```bash
llama-server -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF:Q4_K_M --jinja
# or the gated BF16 weights, after accepting the card's conditions:
vllm serve bottlecapai/ThinkingCap-Qwen3.8-27B --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml
```

```bash
curl http://localhost:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages": [{"role": "user", "content": "Hello!"}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.5/5, desk reviews from public material, no calls made)

- ★★★☆☆ A gated history hides the 6 October change (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ PolyForm Small Business and a contact-details gate (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Reasoning fields named per runtime (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Gated hashes and a trust centre without a flaw route (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings from 2 to 3. They agree the gate on the main repository hides its checksums and commit history, that the card documents reasoning_effort and where the reasoning returns, and that PolyForm Small Business limits free commercial use to small companies. Read it as a Qwen3.8-27B with shorter reasoning and many BottleCap builds for a 24 GB card, benchmarked only by BottleCap.