Instructions to use TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4 with Transformers:

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

pipe = pipeline("text-generation", model="TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4")
model = AutoModelForCausalLM.from_pretrained("TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4", 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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "TentaFlow/Bielik-4.5B-v3.0-Instruct-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": "TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4

SGLang

How to use TentaFlow/Bielik-4.5B-v3.0-Instruct-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 "TentaFlow/Bielik-4.5B-v3.0-Instruct-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": "TentaFlow/Bielik-4.5B-v3.0-Instruct-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 "TentaFlow/Bielik-4.5B-v3.0-Instruct-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": "TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4 with Docker Model Runner:
```
docker model run hf.co/TentaFlow/Bielik-4.5B-v3.0-Instruct-NVFP4
```

Bielik-4.5B-v3.0-Instruct-NVFP4

File size: 1,439 Bytes

57d8b5d

---
license: apache-2.0
language:
- pl
base_model:
- speakleash/Bielik-4.5B-v3.0-Instruct
pipeline_tag: text-generation
library_name: transformers
tags:
- bielik
- polish
- nvfp4
- fp4
- compressed-tensors
- vllm
- quantized
---

# Bielik-4.5B-v3.0-Instruct — NVFP4 (W4A4)

Kwantyzacja **NVFP4 (W4A4)** modelu [speakleash/Bielik-4.5B-v3.0-Instruct](https://huggingface.co/speakleash/Bielik-4.5B-v3.0-Instruct),
w formacie `compressed-tensors` (`nvfp4-pack-quantized`) gotowym do **vLLM**.

- **Wagi:** 4-bit (FP4 E2M1), grupy po 16, blokowe skale FP8 (E4M3) + globalna skala FP32
- **Aktywacje:** 4-bit FP4 (dynamiczne, per-grupa)
- **Pominięte:** `lm_head` (pełna precyzja)
- **Rozmiar:** ~8,9 GB (bf16) → **~2,7 GB**

## Kalibracja

PTQ przez [`llm-compressor`](https://github.com/vllm-project/llm-compressor), 512 próbek polskiej Wikipedii,
długość sekwencji 2048. Kalibracja na polskim tekście dla zachowania jakości w języku docelowym.

## Sprzęt

NVFP4 jest akcelerowany sprzętowo na **Blackwellu (sm_100+)**. Na starszych GPU (np. Ada/RTX 4090) vLLM
ładuje model przez kernel **Marlin (weight-only FP4)** — działa wszędzie, zysk to mniejszy VRAM.

## Użycie (vLLM)

```bash
vllm serve <repo>/Bielik-4.5B-v3.0-Instruct-NVFP4 --max-model-len 8192
```

## Licencja i atrybucja

Apache-2.0, dziedziczona z modelu bazowego (SpeakLeash & ACK Cyfronet AGH).
Ten checkpoint to wyłącznie kwantyzacja oryginalnych wag.