Instructions to use Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF 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 Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16
# Run inference directly in the terminal:
./build/bin/llama cli -hf Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

Use Docker

docker model run hf.co/Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

Llama new
LM Studio
Jan

vLLM

How to use Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF",
		"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/Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

Ollama
How to use Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF with Ollama:
```
ollama run hf.co/Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16
```
Unsloth Desktop

How to use Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

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": "Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF with Docker Model Runner:
```
docker model run hf.co/Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16
```

Lemonade

How to use Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

Run and chat with the model

lemonade run user.ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF-BF16

List all available models

lemonade list

Hermes Agent

How to use Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

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 Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16

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 "Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF:BF16" \
  --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"

ThinkingCap-Qwen3.8-Flash-Next Mixed-Quant GGUF

Independent conversion of bottlecapai/ThinkingCap-Qwen3.8-Flash-Next, BottleCap AI's shorter-thinking post-training of Qwen3.8-Flash-Next. Source revision: d90c9fedadaf7556df4206e6bea5a6b391b1a58e. BottleCap AI's thinking-length and accuracy results were measured on the BF16 checkpoint with vLLM; see the source card. They were not re-measured for this quantization.

The tensor precision map matches the Swift1.5 and Darwin-180B-RSI Q5 packages. Main weights are converted directly from source BF16, without imatrix calibration, pruning, expert dropping, layer dropping or requantization from another GGUF.

Storage

Artifact	Files	Size
Main GGUF / 1,628 tensors	3	77.5559 GiB
Official FP8 SSD-PLE / 128 parts	4	47.6841 GiB
Total weights, including BF16 scale	7 + scale	125.2400 GiB

The 51.2B-parameter PLE table stays on SSD with a bounded runtime cache. The 128.8B-parameter compute backbone is GPU resident. Active KV, workspaces, derived alignment artifacts and headroom are additional memory costs.

Precision map

Region	Type
Routed gate/up, layers 2–45	Q4_K
Routed gate/up, layers 0, 1, 46, 47	Q5_K
Routed down, main 512 columns	Q5_K
Routed down, 128-column tail	Q5_0
MTP and most always-active matrices	Primarily Q8_0
Remaining matrices, norms and controls	BF16 / F32 / I64
PLE n-gram table	FP8 E4M3FN, shared BF16 scale `0x3951`

The tensor map retains vision, embedded MTP and PLE controls. Source tokenizer, processors and Jinja are included; config, tokenizer and template are byte-identical to Qwen. The Rust host preserves source NFC normalization and its Unicode-mark-aware pre-tokenizer.

PLE reuse

All 128 ThinkingCap BF16 PLE parts match the pinned Qwen BF16 reference. Complete source-shard SHA-256 proves 124 parts; direct tensor-range SHA-256 proves the four parts in the two changed mixed shards. See the identity report.

After that gate, the official Qwen FP8 PLE was reused unchanged: the same four files as the Baekpica Qwen, Swift and Darwin packages. The fresh file hashes are in SHA256SUMS.

Thinking control

Thinking is on unless a request turns it off. ds4-dfm-rs renders the source template's reasoning instructions from the request's effort:

Request effort	Template instruction
`high`, `xhigh`, `max`	`xhigh`, the source template's default
omitted, `low`, `medium`, `minimal`	`low`
`none`, `off`, or Anthropic `"thinking": {"type": "disabled"}`	thinking off

An omitted effort takes ds4-dfm-rs's low default, not the template's xhigh. BottleCap AI's published results use xhigh; send "reasoning_effort": "high" for that setting. The template's instruction-free medium mode is not exposed.

Performance reference

These measurements and the graph are from Qwen Base Q5 on one DGX Spark / GB10. ThinkingCap uses the same Q5 recipe and qwen4exp execution path; its throughput has not been measured. The FP8 row is the reference for this package's PLE format.

Qwen Base Q5 PLE	Prefill tok/s	Decode tok/s
BF16	1,245.1	28.59
FP8	1,323.1	28.93

The table uses the median of three run means across 32 frontiers, from 2K to 64K. Each process keeps one warm session, appends 2,048 input tokens and generates 128 greedy tokens per frontier. Both formats use a 2 GiB PLE cache, 16 workers and MTP draft 2. Curves show per-frontier medians; bands show observed ranges. One BF16 run disabled MTP automatically and remains in the results. Decode differences include generated-token and draft-policy effects.

Original Qwen card · Protocol and raw data · Graph and table provenance

Run

Use the Rust host in ds4-dfm-rs at 8105f4ea or newer, which accepts this source revision. ThinkingCap input requires the Rust hosts. Standalone C hosts reject its declared NFC/Unicode tokenizer at vocabulary loading. See reproduction for build and conversion commands.

tc_root=./ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF
hf download Baekpica/ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant-GGUF --local-dir "$tc_root"
(cd "$tc_root" && sha256sum -c SHA256SUMS)
export DS4_QWEN_PLE_DIR="$(realpath "$tc_root/PLE-FP8")"
export DS4_QWEN_PLE_CACHE_MB=2048
export DS4_QWEN_PLE_WORKERS=16
./ds4-server --cuda -m "$tc_root/MQ-Q5-SSD-PLE-BF16/ThinkingCap-Qwen3.8-Flash-Next-MQ-Q5-SSD-PLE-BF16-00001-of-00003.gguf" \
  -c 8192 --max-seqs 2 --native-chunk 512 --mem-floor-gb 12 \
  --prefix-reuse partial --mtp-mode on --mtp-draft 2 \
  --kv-disk-dir ./thinkingcap-kv --kv-disk-space 32G \
  --host 127.0.0.1 --port 8000 --model-id ThinkingCap-Qwen3.8-Flash-Next-Mixed-Quant \
  --no-update-check

Set DS4_QWEN_PLE_DIR explicitly. MQ-Q5-SSD-PLE-BF16 is the inherited main packaging name; this repository bundles the FP8 PLE sidecar.

Verification

Source-file digests, all 1,658 source tensors, exact 1,628-tensor Q5 recipe, PLE controls/layout, full BF16 PLE identity, GGUF structure and independent weight hashes passed. See artifact-manifest.json.

DGX Spark / GB10 CUDA serving passed with build 8105f4ea, an 8,192-token context, 2 banks, native chunk 512, MTP draft 2, a 2 GiB FP8 PLE cache and 32 GiB disk KV. The receipt covers 28 text/image/tool/stream/concurrency API requests, live append/edit/fork reuse, fresh-process disk restoration, cold-response parity, the mandatory 10-token MTP copy probe with speculation_active=true, MTP off, and four effort settings (omitted, high and low with closed reasoning, none without) that each returned the exact product. Fault and shed counters stayed zero. The longest lifecycle prompt was 743 tokens; filled 8K/262K contexts, throughput and quantized benchmark quality were not measured. Raw-input tokenization matched the pinned source tokenizer on 7 recorded vectors. Public unauthenticated verification of all seven weight hashes/sizes and complete reads of every small file is recorded in the publication receipt. Original-model benchmark scores do not establish this quantization's quality.

License

Qwen Community License 1.0, Copyright 2026 Qwen. ThinkingCap is released by BottleCap AI under the same license. NOTICE records conversion changes. The quantizer is MIT licensed; see ds4/LICENSE. Runtime: antirez/ds4, Entrpi/ds4 and the ggml-org/llama.cpp quant formats. Independent Baekpica conversion, not an official BottleCap AI or Qwen release.