Instructions to use bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16
# Run inference directly in the terminal:
llama cli -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16
# Run inference directly in the terminal:
llama cli -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16
# Run inference directly in the terminal:
./llama-cli -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Use Docker

docker model run hf.co/bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

LM Studio
Jan

vLLM

How to use bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Ollama
How to use bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw with Ollama:
```
ollama run hf.co/bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16
```
Unsloth Desktop

How to use bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw with Docker Model Runner:
```
docker model run hf.co/bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16
```

Lemonade

How to use bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Run and chat with the model

lemonade run user.ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw-F16

List all available models

lemonade list

Hermes Agent

How to use bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw:F16" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw

A 14.3 GB GGUF of bottlecapai/ThinkingCap-Qwen3.8-27B that runs on an 18 GB GPU with image input, MTP speculative decoding and a 32k-token context all on the card, and matched the bf16 weights on all five benchmarks we evaluated (2,863 questions: 81.00% against 81.03%).

The standard GGUFs (ThinkingCap-Qwen3.8-27B-GGUF) start at IQ4_XS, 15.5 GB, which with the vision projector and a 32k context does not fit 18 GB.

Files

File	What	Size
`ThinkingCap-Qwen3.8-27B-RCO-4.2bpw.gguf`	language model + MTP head, mixed precision (4.19 bits per weight on average)	14.3 GB
`mmproj-ThinkingCap-Qwen3.8-27B-f16.gguf`	vision projector (mmproj), f16	931 MB

How the precision is chosen

Every weight tensor was quantized with llama.cpp's own types (IQ2_S up to Q8_0, all with one importance matrix computed on the model's own reasoning traces). Which type each of the 401 language-model weight matrices gets was then picked by an RCO search: it minimizes the KL divergence between the bf16 model's and the mixed model's next-token distributions on 256 of the model's own traces, under an exact file-size budget. The same idea as ISTA-DASLab's GSQ-RCO GGUFs, searched on this finetune rather than the base model; the rounding itself is llama.cpp's.

part	bits per weight
whole file, all 27.3B weights (bf16 original: 16)	4.19
feed-forward (`ffn_gate` / `ffn_up` / `ffn_down`), most of the weights	3.8–3.9
linear-attention (Gated DeltaNet) projections	3.9–4.4
full-attention layers: q / output / k / v	4.4 / 5.0 / 6.1 / 6.9
output head (`lm_head`)	8.5 (Q8_0)
MTP head	6.6 (Q6_K)
token embedding (a lookup, kept in system RAM by llama.cpp)	2.6 (IQ2_S)

Usage (llama.cpp)

Tested with llama.cpp commit 6b790a9. The configuration measured to fit an 18 GB card:

huggingface-cli download bottlecapai/ThinkingCap-Qwen3.8-27B-GGUF-RCO-4.2bpw ThinkingCap-Qwen3.8-27B-RCO-4.2bpw.gguf mmproj-ThinkingCap-Qwen3.8-27B-f16.gguf --local-dir .

llama-server -m ThinkingCap-Qwen3.8-27B-RCO-4.2bpw.gguf --mmproj mmproj-ThinkingCap-Qwen3.8-27B-f16.gguf \
  -c 32768 -ngl 999 -fa on -ctk q8_0 -ctv q8_0 \
  --spec-type draft-mtp --spec-draft-n-max 3 --jinja

GPU memory of the llama-server process with a prompt filling the whole context, followed by an image request (one slot, flash attention, vision projector and MTP drafter on the GPU):

context	KV cache	GPU memory	left on an 18 GB card (18,432 MiB)
32,768 (command above)	q8_0	17,120 MiB	1,312 MiB
40,960	q8_0	17,472 MiB	960 MiB
49,152	q8_0	17,824 MiB	608 MiB
32,768	f16	17,992 MiB	440 MiB

32k is the tested default and 40k fits with about 1 GB to spare; an f16 KV cache is not recommended on 18 GB.

Use the sampling settings from the main model card (temperature 1.0, top_p 0.95, top_k 20, min_p 0.0). Greedy decoding can loop.

MTP: the file carries the model's multi-token-prediction head; --spec-type draft-mtp drafts with it, no separate draft model. Runtimes without MTP support for this architecture may refuse to load the file (missing tensor blk.64…) — update the runtime.
Vision: --mmproj enables image input (LM Studio and similar apps pick up the mmproj-*.gguf automatically). On a smaller card, --no-mmproj-offload keeps the projector on the CPU at the cost of slower image prompts.

Expected performance

Paired comparison with the bf16 weights served by vLLM 0.29.0 on an RTX PRO 6000 (MTP k=3), on the same questions with the same seed: RealWorldQA 765 questions (images), GPQA-Diamond 198, MMLU-Pro 1,500 (a fixed slice), IFBench 300, AA-LCR 100 (long documents, graded by Gemini 3.1 Flash-Lite with a 1,024-token thinking budget). One seed each, thinking at the chat template's default reasoning effort (xhigh), sampled decoding (temperature 1.0, top_p 0.95, top_k 20, min_p 0.0). This file ran on llama.cpp (6b790a9, --jinja, vision projector loaded) on the same GPU type, one per benchmark, without MTP.

benchmark (questions)	accuracy %, bf16 → this file	Δ accuracy, pp [95% CI] (McNemar p)	tokens mean / median / p95, bf16 → this file	Δ mean tokens [95% CI]	Δ median tokens
RealWorldQA (765)	82.1 → 82.6	+0.5 [−1.7, +2.7] (p 0.724)	673 / 112 / 3,569 → 531 / 126 / 2,908	−21.1% [−37.2, +1.1]	+12.5%
GPQA-Diamond (198)	87.9 → 87.4	−0.5 [−4.1, +3.1] (p 1.000)	6,420 / 1,076 / 35,054 → 6,494 / 1,053 / 29,947	+1.1% [−14.6, +20.4]	−2.1%
MMLU-Pro (1,500)	79.9 → 79.7	−0.2 [−1.8, +1.4] (p 0.866)	2,094 / 314 / 11,365 → 1,991 / 290 / 11,822	−4.9% [−15.8, +7.1]	−7.9%
IFBench (300)	79.0 → 78.3	−0.7 [−5.0, +3.7] (p 0.880)	4,339 / 1,758 / 18,498 → 4,435 / 1,734 / 19,038	+2.2% [−10.0, +14.9]	−1.4%
AA-LCR (100)	83.0 → 84.0	+1.0 [−4.2, +6.2] (p 1.000)	1,758 / 964 / 5,287 → 1,594 / 928 / 5,843	−9.3% [−25.5, +13.6]	−3.7%
all five (2,863)	81.0 → 81.0	−1 question

Δ accuracy: this file minus bf16 on the same questions, with a 95% interval over questions and the exact McNemar p. Tokens are completion tokens (reasoning plus answer); the mean-token interval is a bootstrap over questions. No accuracy or length difference is detected on any benchmark.

Decode speed and MTP self-speculative decoding (MMLU-Pro) — 24 questions × 1 seed, llama.cpp 6b790a9, 4 parallel slots, MTP drafting 3 tokens

GPU	config	median tokens	tok/s	MTP speedup	accept_len (max 4)
RTX 4090	standard	236	36.1	1.00×	—
RTX 4090	MTP	166	47.8	1.32×	2.51
H200	standard	241	40.9	1.00×	—
H200	MTP	316	59.5	1.45×	2.46

Where to find us

Need even more efficiency? The open release is production-ready. Our enterprise versions go further — fewer thinking tokens still, tuned to your workload, at matched accuracy on your own tasks. Built for AI labs, inference providers and enterprises running models at scale. Deployed on your infrastructure, or in the cloud and region you choose. Talk to our team

License

ThinkingCap: PolyForm Small Business 1.0.0 + BottleCap personal-use grant (see LICENSE).

Upstream Qwen materials: Apache-2.0 (see NOTICE).

Commercial license: contact BottleCap AI.

Citation

If you use this model, please cite:

@misc{ThinkingCap-Qwen3.8-27B,
  title     = {bottlecapai/ThinkingCap-Qwen3.8-27B},
  author    = {Osusky, Adam and Lindauer, Jan and Jirkovsky, Adam and Mihal, Filip and Platek, Ondrej and Herel, David and Ihnatchenko, Luka and Bartek, Vojtech and Jirak, Jiri and Kubista, Daniel and Krus, Frantisek and Mikolov, Tomas},
  year      = {2026},
}