Instructions to use Ternweft/Kwen1.0-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Ternweft/Kwen1.0-4B with Transformers:

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

pipe = pipeline("image-text-to-text", model="Ternweft/Kwen1.0-4B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Ternweft/Kwen1.0-4B")
model = AutoModelForMultimodalLM.from_pretrained("Ternweft/Kwen1.0-4B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Ternweft/Kwen1.0-4B 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 Ternweft/Kwen1.0-4B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Ternweft/Kwen1.0-4B:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Ternweft/Kwen1.0-4B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Ternweft/Kwen1.0-4B:Q4_K_M

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 Ternweft/Kwen1.0-4B:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf Ternweft/Kwen1.0-4B:Q4_K_M

Use Docker

docker model run hf.co/Ternweft/Kwen1.0-4B:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use Ternweft/Kwen1.0-4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Ternweft/Kwen1.0-4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ternweft/Kwen1.0-4B",
		"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/Ternweft/Kwen1.0-4B:Q4_K_M

SGLang

How to use Ternweft/Kwen1.0-4B 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 "Ternweft/Kwen1.0-4B" \
    --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": "Ternweft/Kwen1.0-4B",
		"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 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 "Ternweft/Kwen1.0-4B" \
        --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": "Ternweft/Kwen1.0-4B",
		"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"
						}
					}
				]
			}
		]
	}'

Ollama
How to use Ternweft/Kwen1.0-4B with Ollama:
```
ollama run hf.co/Ternweft/Kwen1.0-4B:Q4_K_M
```
Unsloth Desktop

How to use Ternweft/Kwen1.0-4B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Ternweft/Kwen1.0-4B:Q4_K_M

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": "Ternweft/Kwen1.0-4B:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Ternweft/Kwen1.0-4B with Docker Model Runner:
```
docker model run hf.co/Ternweft/Kwen1.0-4B:Q4_K_M
```

Lemonade

How to use Ternweft/Kwen1.0-4B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Ternweft/Kwen1.0-4B:Q4_K_M

Run and chat with the model

lemonade run user.Kwen1.0-4B-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Ternweft/Kwen1.0-4B with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Ternweft/Kwen1.0-4B:Q4_K_M

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 Ternweft/Kwen1.0-4B:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Ternweft/Kwen1.0-4B with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Ternweft/Kwen1.0-4B:Q4_K_M

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 "Ternweft/Kwen1.0-4B:Q4_K_M" \
  --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"

Kwen1.0-4B — retention checkpoint

This repository now contains the new 100-step retention checkpoint, replacing the earlier 700-step pilot at the owner's request. The repository URL and GGUF filename remain stable. This is an experimental release: the new checkpoint failed its frozen acceptance criteria and is not a validated company starter model or a general upgrade over Qwen3.5-4B. No 9B comparison was performed.

The previous release, including its weights, GGUF and original results, is preserved at v1-original-before-retention (commit 35e325aff96a261d60c0b7ec64c713db94eb7fba). Downloads from main now refer to the new checkpoint. Do not compare the percentages below directly with the old release's table: different samples and revised research orchestration were used.

What changed

A fresh attention-only LoRA adapter starts from pinned original Qwen3.5-4B (851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a). It uses rank 16, alpha 32, dropout 0.05, learning rate 1e-5, and 100 fixed optimizer steps. The objective combines supervised tool examples with forward KL from the original model on coding/general replay contexts. There are 400 tool and 400 replay conversations; 12,386,304 parameters were trainable. The checkpoint was fixed in advance, not selected by final benchmark results.

Training sources are NousResearch/hermes-function-calling-v1, Team-ACE/ToolACE, HuggingFaceTB/smoltalk (smol-summarize and smol-constraints), and m-a-p/CodeFeedback-Filtered-Instruction. Source licenses are Apache-2.0; see the attached original data provenance and retention manifest. Retain upstream notices. This project is not affiliated with Qwen or Nous Research.

Measured results

Development suite	Cases	Original base	New checkpoint
mbpp	50	78%	74%
mmlu_pro	50	46%	46%
ifeval	50	80%	78%
research	10	10%	0%
tools	25	40%	52%

These are sampled deterministic non-thinking results for merged BF16 weights. Tools improved, while coding, instruction following and research declined. Both arms passed four of six manually reviewed companion dialogues; both failed the API-research and factual-correction cases. The original frozen gates remain failed; the owner's publication request does not change that result. No independent final validation was run. See RESULTS.md, evaluation.json, and companion-review.json for evidence and uncertainty.

Files and use

Root: merged BF16 Transformers weights, including unvalidated base vision components.
adapter/: the new retention adapter and native tokenizer. Load it only on the pinned original base, not on these already-merged weights.
gguf/Kwen1.0-4B-Q4_K_M.gguf: Q4_K_M text-only quantization of this checkpoint. It has a separate three-prompt CPU load/chat/code/tool API smoke check in quantized-smoke.json. This is not a full quantized benchmark or an RTX 5070 speed test. The old release's 12-case quantization report does not apply.

Keep the native Qwen3.5 template. The reported development results use non-thinking generation with an 8,192-token context allowance. Vision and thinking-mode behavior of this checkpoint are unvalidated. Optional MTP heads are omitted; standard autoregressive decoding is supported. The separate canceled original-base thinking-profile experiment does not evaluate these weights.

The prior release can be loaded by specifying revision="v1-original-before-retention". Model quality limitations remain; public benchmark pretraining exposure cannot be excluded.

Downloads last month: 135

Safetensors

Model size

5B params

Tensor type

BF16

Model tree for Ternweft/Kwen1.0-4B

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Finetuned

(1,384)

this model