Instructions to use yet-another-ai/Jet-Qwen3.5-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use yet-another-ai/Jet-Qwen3.5-4B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="yet-another-ai/Jet-Qwen3.5-4B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("yet-another-ai/Jet-Qwen3.5-4B")
model = AutoModelForCausalLM.from_pretrained("yet-another-ai/Jet-Qwen3.5-4B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

PEFT
How to use yet-another-ai/Jet-Qwen3.5-4B with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use yet-another-ai/Jet-Qwen3.5-4B 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 yet-another-ai/Jet-Qwen3.5-4B:Q8_0
# Run inference directly in the terminal:
llama cli -hf yet-another-ai/Jet-Qwen3.5-4B:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf yet-another-ai/Jet-Qwen3.5-4B:Q8_0
# Run inference directly in the terminal:
llama cli -hf yet-another-ai/Jet-Qwen3.5-4B:Q8_0

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 yet-another-ai/Jet-Qwen3.5-4B:Q8_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf yet-another-ai/Jet-Qwen3.5-4B:Q8_0

Use Docker

docker model run hf.co/yet-another-ai/Jet-Qwen3.5-4B:Q8_0

Llama new
LM Studio
Jan

vLLM

How to use yet-another-ai/Jet-Qwen3.5-4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "yet-another-ai/Jet-Qwen3.5-4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "yet-another-ai/Jet-Qwen3.5-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/yet-another-ai/Jet-Qwen3.5-4B:Q8_0

SGLang

How to use yet-another-ai/Jet-Qwen3.5-4B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "yet-another-ai/Jet-Qwen3.5-4B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "yet-another-ai/Jet-Qwen3.5-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "yet-another-ai/Jet-Qwen3.5-4B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "yet-another-ai/Jet-Qwen3.5-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use yet-another-ai/Jet-Qwen3.5-4B with Ollama:
```
ollama run hf.co/yet-another-ai/Jet-Qwen3.5-4B:Q8_0
```
Unsloth Desktop

How to use yet-another-ai/Jet-Qwen3.5-4B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf yet-another-ai/Jet-Qwen3.5-4B:Q8_0

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": "yet-another-ai/Jet-Qwen3.5-4B:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use yet-another-ai/Jet-Qwen3.5-4B with Docker Model Runner:
```
docker model run hf.co/yet-another-ai/Jet-Qwen3.5-4B:Q8_0
```

Lemonade

How to use yet-another-ai/Jet-Qwen3.5-4B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull yet-another-ai/Jet-Qwen3.5-4B:Q8_0

Run and chat with the model

lemonade run user.Jet-Qwen3.5-4B-Q8_0

List all available models

lemonade list

Hermes Agent

How to use yet-another-ai/Jet-Qwen3.5-4B with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf yet-another-ai/Jet-Qwen3.5-4B:Q8_0

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 yet-another-ai/Jet-Qwen3.5-4B:Q8_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use yet-another-ai/Jet-Qwen3.5-4B with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf yet-another-ai/Jet-Qwen3.5-4B:Q8_0

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 "yet-another-ai/Jet-Qwen3.5-4B:Q8_0" \
  --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"

Jet Qwen3.5 4B

A text-only Qwen3.5 model adapted for Jet, a local structured decision engine. Jet scores a fixed set of semantic answers with teacher forcing and returns choices and candidate probabilities.

Try the WebGPU playground · Jet documentation

Paper and citation

The JET method is described in JET: Justification Evaluation in Transformer by Shenghao Ding (2026), arXiv:2609.33874. If you use JET or the artifacts in this repository, please cite:

@misc{ding2026jet,
  title = {{JET}: Justification Evaluation in Transformer},
  author = {Shenghao Ding},
  year = {2026},
  eprint = {2609.33874},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url = {https://arxiv.org/abs/2609.33874}
}

What is released

Root: merged BF16 Transformers text-model weights and tokenizer.
adapter/: original PEFT LoRA adapter and tokenizer; load with subfolder="adapter".
gguf/Jet-Qwen3.5-4B-Q8_0.gguf: merged Q8_0 GGUF for Jet/llama.cpp.
training/: configuration, completion record, export provenance and dataset revisions.
evaluation/: archived native evaluation reports.
SHA256SUMS: checksums of all published artifacts except this checksum file itself.

The base is Qwen/Qwen3.5-4B, revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. This release contains the text model only, without a vision tower or projector. It does not provide image inference by itself.

Use with Jet

Download gguf/Jet-Qwen3.5-4B-Q8_0.gguf from this repository, then run:

jet serve --model-path Jet-Qwen3.5-4B-Q8_0.gguf --model-id yet-another-ai/Jet-Qwen3.5-4B

Send a request to http://127.0.0.1:8080/api/alpha/decisions:

{"model":"yet-another-ai/Jet-Qwen3.5-4B","state":{"ticket":"Login fails for all users"},"questions":{"owner":{"type":"choice","instructions":"Who should own this incident?","criteria":{"app":"Application team","infra":"Infrastructure team"}}}}

Use Jet's prompt renderer and candidate-scoring path to reproduce the intended decision behavior. Loading the model in a generic chat widget does not reproduce Jet scoring. The small browser demo defaults to 0.8B; the 4B Q8_0 model requires native Jet.

Training

Supervised BF16 LoRA fine-tuning on correct quoted semantic answers, using Jet's actual prompt renderer. Only target-continuation tokens contribute to the loss. There is no teacher-model distribution or reasoning-trace distillation in this recipe. It does not train JSON output, answer letters, a classifier head, or image inputs.

One epoch, rank 16, alpha 32, dropout 0.05, learning rate 1e-4, effective batch 16 (microbatch 4 × accumulation 4), seed 42, maximum sequence length 2,048. The archived configuration defaults may differ from launch overrides; the actual batch settings and input hashes are recorded in training/completed.json.

Data preparation used 22,791 training records: BoolQ 9,427, ARC Easy 2,248, ARC Challenge 1,116, and SNLI 10,000. Validation uses a deterministic subset of at most 512 records for checkpoint-loss monitoring. MMLU is held out. The training/export implementation lives in Jet's training directory. Recorded local paths in provenance are normalized to jet for portability; artifact and input hashes are retained.

Evaluation

The original native CUDA evaluation reports are included under evaluation/. These are existing experiment results, not a new evaluation of this upload. They measure Jet candidate scoring with mean-token normalization and Q8_0 weights; they are not chat-generation benchmark scores or calibrated probabilities.

The archived 4B paired MMLU report contains 10,446 adapted-model correct answers out of 14,042 (74.39%). The baseline report records 9,332/14,042. Unlike the 0.8B experiment, this archive does not establish a matched base export/conversion; do not interpret it as a controlled training-effect estimate.

Limits

Training focuses on English text decisions from a small task mixture. Performance on arbitrary domains and languages is not established. Confidence is a relative softmax over the supplied candidates and is not calibrated certainty. Candidate wording and ordering can affect results. Existing held-out overlap checks do not establish absence of contamination in the pretrained base model. Browser WebGPU compatibility and memory limits differ from native CUDA inference.

License and attribution

Model weights and Jet code are released under Apache-2.0; see LICENSE and the original Qwen BASE_MODEL_LICENSE. The base model is by the Qwen team. Training data attribution and source revisions are retained in training/datasets.lock.json: BoolQ (Clark et al., 2019, CC BY-SA 3.0), ARC (Clark et al., 2018, CC BY-SA 4.0), and SNLI (Bowman et al., 2015, CC BY-SA 4.0). No training dataset records are redistributed in this model repository. The dataset licenses remain applicable to the source data.