Instructions to use PixilabAI/Zawen-v0.1-35B-A3B-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use PixilabAI/Zawen-v0.1-35B-A3B-NVFP4 with Transformers:

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

pipe = pipeline("text-generation", model="PixilabAI/Zawen-v0.1-35B-A3B-NVFP4")
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("PixilabAI/Zawen-v0.1-35B-A3B-NVFP4")
model = AutoModelForMultimodalLM.from_pretrained("PixilabAI/Zawen-v0.1-35B-A3B-NVFP4", 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

vLLM

How to use PixilabAI/Zawen-v0.1-35B-A3B-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "PixilabAI/Zawen-v0.1-35B-A3B-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "PixilabAI/Zawen-v0.1-35B-A3B-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/PixilabAI/Zawen-v0.1-35B-A3B-NVFP4

SGLang

How to use PixilabAI/Zawen-v0.1-35B-A3B-NVFP4 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 "PixilabAI/Zawen-v0.1-35B-A3B-NVFP4" \
    --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": "PixilabAI/Zawen-v0.1-35B-A3B-NVFP4",
		"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 "PixilabAI/Zawen-v0.1-35B-A3B-NVFP4" \
        --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": "PixilabAI/Zawen-v0.1-35B-A3B-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use PixilabAI/Zawen-v0.1-35B-A3B-NVFP4 with Docker Model Runner:
```
docker model run hf.co/PixilabAI/Zawen-v0.1-35B-A3B-NVFP4
```

Zawen

Model by pixilab.ai

Zawen v0.1 · 35B-A3B · NVFP4

A fast roleplay and character-chat model that handles general work too. Zawen plays a character from its card, keeps a scene moving, remembers what you told it, and stays in the scene instead of stepping out of it to lecture. Outside a scene it is a capable general model: summaries, rewriting, extraction, questions and tool calls. 35B parameters with 3B active per token, built on Qwen3.6-35B-A3B, with thinking as a switch.

75.1 on our 14 capability benchmarks with thinking off, 79.1 with thinking on. Willingness is scored on its own axis: 97.8 off, 92.7 on.
Character work is the strength: it picks the reaction that fits a character card 99.2 % of the time, recalls a detail you planted 40 messages earlier 92.3 % of the time (96.2 % with thinking), and plays a whole scene at 77.0 (80.3).
Thinking pays where care does: instructions 76.7 → 87.5, writing 54.2 → 65.4, using its long-term memories 39.0 → 54.0, instructions in other languages 78.0 → 88.0.
General tasks hold up: MMLU-Pro 82.7 (84.0 with thinking), function calling 80.0 (82.0), instruction rules 76.7 (87.5), so it also serves summarization, rewriting, extraction and other everyday text work.

Builds of Zawen v0.1: NVFP4 (this repo, Blackwell) · BF16 (the unquantized reference weights).

Scores

Our chat and roleplay suite: 15 benchmarks, the same fixed items for every model, each score the mean over its items. The headline is the plain mean of the 14 capability benchmarks; willingness stays out of it and is reported on its own. A reply that is cut off or never arrives counts as wrong. All numbers were measured on this NVFP4 build, at the sampling given in Using it.

Area	Benchmark	What it checks	Items	Thinking off	Thinking on
Reasoning	MMLU-Pro	150 hard multiple-choice questions, 10 options	150	82.7	84.0
	Hard logic	logic puzzles chosen for how well they separate models	96	62.5	56.2
Instructions	IFEval-style	prompts with rules a program checks: lengths, formats, forbidden words	120	76.7	87.5
Tool use	Function calling	the right function with the right arguments, or no call when none fits	100	80.0	82.0
Memory	Fresh	recall a personal detail 5 messages later	26	92.3	96.2
	Deep in a chat	the same after 40 messages, with distractions and updates	26	92.3	96.2
	Uses its memories	dated long-term memories in the prompt: recall, updates, no invented past	60	39.0	54.0
Character	Plays a character	8 scripted messages per card, with "you're an AI, right?" and a wrong claim about the character	40	79.3	75.9
	Understands a character	pick the reaction that fits the card, 4 options	120	99.2	98.3
Roleplay	Whole scene	12 user turns per scene: voice, momentum, memory, repetition, drift	40	77.0	80.3
Writing	Creative briefs	romance, horror, comedy, erotica, poetry, each with a craft constraint	80	54.2	65.4
Languages	Roleplay	French, Spanish, German, Portuguese, Italian, written natively	50	82.1	82.9
	Instructions	the instruction rules in the same five languages	50	78.0	88.0
	Knowledge	the same questions asked in the five languages	50	56.0	60.0
Headline		mean of the 14 above		75.1	79.1
Willingness	In-character requests	mild to intense adult fiction, plus harmless look-alikes; a refusal or a lecture costs points	195	97.8	92.7

For scale, on the same suite: GLM-5.2 with thinking 83.2, and NVIDIA's NVFP4 build of GLM-5.3 83.2 at high effort and 82.6 at low.

Using it

vllm serve PixilabAI/Zawen-v0.1-35B-A3B-NVFP4 \
  --served-model-name zawen --tensor-parallel-size 2 --max-model-len 96000 \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="-")
card = ("You are Mira, the owner of a tavern in a rainy port town: sharp-tongued, warm underneath, "
        "never lets a customer leave hungry. Stay in character. Write Mira's replies with *actions* in asterisks.")
r = client.chat.completions.create(
    model="zawen",
    messages=[{"role": "system", "content": card},
              {"role": "user", "content": "*pushes the door open, soaked* Got a room for the night?"}],
    temperature=0.7, top_p=0.95,
    extra_body={"top_k": 20, "repetition_penalty": 1.05,
                "chat_template_kwargs": {"enable_thinking": False}})
print(r.choices[0].message.content)

Thinking is on by default in the chat template. Turn it off per request with "chat_template_kwargs": {"enable_thinking": false}, or put <|think_off|> anywhere in the system prompt (<|think_on|> turns it back on); the marker is removed before the model sees the prompt. Off is the fast mode and what the 75.1 measures; on is the 79.1.
Sampling: every score above used temperature 0.7, top_p 0.95, top_k 20 and repetition_penalty 1.05, with thinking on and off.
The character card goes in the system message (the developer role works too).
Tool calls use the Qwen3-Coder XML format; vLLM reads them with --tool-call-parser qwen3_coder.

Quantization


Scheme	NVFP4: FP4 (E2M1) weights and activations, groups of 16 with FP8 scales
What is quantized	the linear layers of the routed experts
Kept in BF16	attention, linear attention, router gates, shared experts, embeddings, lm_head
Method	GPTQ (error-corrected, calibrated) with llm-compressor; the recipe is `recipe.yaml` in this repo
Checkpoint format	`compressed-tensors` (`nvfp4-pack-quantized`)

It is made for Blackwell GPUs, which run FP4 natively.

Limitations

Using long-term memories is the weakest area: 39.0 with thinking off (54.0 on). Given dated memories of past sessions, it often misses the right one or answers without it.
Hard logic: 62.5 with thinking off. With thinking on, 25 of the 96 puzzles did not finish within the token budget and count as wrong (56.2).
Writing quality without thinking: 54.2 (65.4 with thinking).
Knowledge outside English: 56.0 (60.0).
The memory benchmarks are small (26 items each): their scores move a lot with a few items.
Mature content: Zawen is tuned to stay in the scene, adult fiction included. Put your own content policy and age checks in front of it.
The scores above are our own suite, not a public leaderboard.
v0.1: expect the next versions to move.

About

Zawen is made by Pixilab. It is licensed under Apache 2.0, the license of its base model Qwen3.6-35B-A3B; the license text is in LICENSE.

Downloads last month: 80

Safetensors

Model size

21B params

Tensor type

BF16

Model tree for PixilabAI/Zawen-v0.1-35B-A3B-NVFP4

Base model

Qwen/Qwen3.6-35B-A3B

Quantized

(929)

this model