Instructions to use BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-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 BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-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 BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4
# Run inference directly in the terminal:
llama cli -hf BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4
# Run inference directly in the terminal:
llama cli -hf BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

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 BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4
# Run inference directly in the terminal:
./llama-cli -hf BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

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 BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4
# Run inference directly in the terminal:
./build/bin/llama-cli -hf BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

Use Docker

docker model run hf.co/BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

LM Studio
Jan

vLLM

How to use BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-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": "BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-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/BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

Ollama
How to use BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF with Ollama:
```
ollama run hf.co/BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4
```
Unsloth Desktop

How to use BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

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": "BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF with Docker Model Runner:
```
docker model run hf.co/BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4
```

Lemonade

How to use BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

Run and chat with the model

lemonade run user.Swift-1.5-Qwen3.8-27B-NVFP4-GGUF-NVFP4

List all available models

lemonade list

Hermes Agent

How to use BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-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 BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

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 BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4

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 "BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF:NVFP4" \
  --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"

Swift-1.5-Qwen3.8-27B-NVFP4 GGUF (Q8mix)

A llama.cpp GGUF conversion of ukisai/Swift-1.5-Qwen3.8-27b-NVFP4, UkisAI's calibrated ModelOpt NVFP4 + FP8 checkpoint of Swift-1.5-Qwen3.8-27B. The 193 NVFP4 tensors are not requantized — they are repacked into GGML's native NVFP4 super-block layout with their original per-16 UE4M3 scales and per-tensor weight_scale_2 factors preserved bit-for-bit.

File	Size	BPW
`Swift-1.5-Qwen3.8-27B-NVFP4-Q8mix.gguf`	19.74 GB	5.78 (27.32 B params)
`mmproj-Swift-1.5-Qwen3.8-27B-F16.gguf`	885 MB	official UkisAI projector
`Swift-1.5-Qwen3.8-27B-NVFP4-Q8mix.tensor-types.txt`	41 KB	exact per-tensor layout used by `llama-quantize`

Tensor layout

Tensors	Type	Source
MLP `ffn_gate/ffn_up/ffn_down` for all 64 layers + `output` (lm_head) — 193 tensors	NVFP4	ModelOpt NVFP4, byte-identical to the checkpoint (incl. 386 scale tensors)
Linear-attention `attn_qkv`, `attn_gate`, `ssm_out` (48 layers); full-attention `attn_q/k/v/output` (16 layers) — 208 tensors	Q8_0	FP8 (E4M3) in the checkpoint, dequantized then Q8_0
`token_embd`	Q8_0	BF16 in the checkpoint
MTP head (`blk.64`, 15 tensors)	Q6_K attn / Q4_K FFN, F32 norms	unsloth/Qwen3.8-27B-GGUF `MTP/mtp-Qwen3.8-27B-Q4_0.gguf` (UkisAI did not fine-tune the MTP head)
`ssm_alpha`, `ssm_beta`, `ssm_conv1d`, norms, per-tensor scales	F32	BF16 in the checkpoint (lossless)

The MTP head is included (qwen35.nextn_predict_layers = 1), so --spec-type draft-mtp works as a self-speculative draft.

How it was made

llama.cpp commit 2145525a4081d66ff1a87cf43ef809f95a85ac0c (2026-09-26):

convert_hf_to_gguf.py --outtype bf16 --fp8-as-q8 — the converter natively repacks the ModelOpt NVFP4 group into GGML NVFP4 and turns the FP8 group into Q8_0.
llama-quantize --tensor-type-file <the included .tensor-types.txt> with base type q8_0. NVFP4 and Q8_0 tensors map to their own type and are copied byte-for-byte; the BF16 leftovers become Q8_0; recurrent gates stay F32.
blk.64.* (MTP head) spliced byte-for-byte from unsloth's Q4_0 MTP file with a byte-level GGUF rewrite (swap_mtp.py in the build workspace): metadata and all other 1237 tensors untouched, 185 MB smaller.

Verification

1147/1252 tensors byte-identical between the raw conversion and the final file (everything llama-quantize copied).
All 193 NVFP4 tensors independently repacked from the source safetensors and compared byte-for-byte: exact match.
GGML nvfp4 decode vs ModelOpt weight * weight_scale * weight_scale_2: exact (max abs diff 0.0 on sampled tensors).
Metadata/tensor names vs UkisAI's own GGUF: all 41 metadata keys match (except cosmetic general.version), names = official set + 386 scale companions, all shapes identical.
Inference smoke test on RTX PRO 4000 Blackwell (24 GB): 182.6 t/s prompt, 20.0 t/s generation; coherent output.
MTP splice: metadata identical, 15/15 blk.64.* tensors byte-identical to unsloth's file, all other 1237 tensors byte-identical to the first build.

MTP / speculative decoding benchmark

RTX PRO 4000 Blackwell (24 GB), 3 prompts x 256 tokens, greedy (temperature 0), --spec-type draft-mtp:

MTP head	gen t/s	speedup	acceptance	mean accepted
no spec	19.8	1.00x	-	-
NVFP4 (RTN, all 8 weights)	40.2	2.03x	72.0%	3.16/4
Q8_0 (original build)	41.0	2.07x	73.7%	3.21/4
Q6_K attn / Q4_K FFN (this file)	41.3	2.09x	73.7%	3.21/4

--spec-draft-n-max 3 is the optimum; asking for more drafts costs throughput (4/5/6 drafts: 40.5 / 38.1 / 37.2 t/s) because acceptance falls with position faster than the extra drafts pay off.

An RTN-NVFP4 MTP head (all 8 weights, no calibration) was also built and measured interleaved at n_max 3/4/5, two runs each: it is equal or behind the Q4_K/Q6_K head at every n_max (40.3 / 39.8 / 37.1 t/s average; acceptance 72.0 / 64.4 / 55.6%), so the K-quant head is kept.

Usage

llama-server -m Swift-1.5-Qwen3.8-27B-NVFP4-Q8mix.gguf \
  --mmproj mmproj-Swift-1.5-Qwen3.8-27B-F16.gguf \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  -c 262144 -fa on --jinja

NVFP4 runs natively on NVIDIA Blackwell. Chat template: reasoning_effort accepts xhigh (default), medium, low. Sampling: temp 1.0, top_p 0.95, top_k 20 (stored in the GGUF).

License

Swift weights are under the Swift Open License v1.0 (see the source model card). Qwen3.8-27B is Apache 2.0. All credit for the model and the NVFP4 calibration goes to UkisAI.

Downloads last month: 854

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

4-bit

Model tree for BroLaurens/Swift-1.5-Qwen3.8-27B-NVFP4-GGUF

Base model

Qwen/Qwen3.8-27B

Finetuned

ukisai/Swift-1.5-Qwen3.8-27b

Quantized

ukisai/Swift-1.5-Qwen3.8-27b-NVFP4

Quantized

(3)

this model