---
language:
  - en
license: apache-2.0
library_name: transformers
tags:
  - mixture-of-experts
  - MoE
  - coding
  - 4-bit
  - nf4
  - bitsandbytes
  - quantized
  - LFM2
  - Qwen
base_model:
  - Akahsizrr/fuse-1-Lite
  - LiquidAI/LFM2.5-2.6B
  - Qwen/Qwen3.6-35B-A3B
pipeline_tag: text-generation
inference: false
---

# fuse-1 Lite — 4-bit Quantized (NF4)

> **4-bit NF4 quantized version of [fuse-1 Lite](https://huggingface.co/Akahsizrr/fuse-1-Lite). 3.36 GB VRAM — runs on T4, RTX 3060, and consumer GPUs.**

This is the bitsandbytes 4-bit quantized version of fuse-1 Lite, using NF4 (NormalFloat 4-bit) quantization with double quantization for maximum compression.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

model = AutoModelForCausalLM.from_pretrained(
    "Akahsizrr/fuse-1-Lite-4bit",
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("Akahsizrr/fuse-1-Lite-4bit")

messages = [{"role": "user", "content": "Write a Python function to check if a number is prime."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.1)

print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
```

## VRAM Requirements

| Precision | VRAM | GPU |
|-----------|------|-----|
| **4-bit (this model)** | **3.36 GB** | T4, RTX 3060, M2 Pro |
| 8-bit | 6.00 GB | T4, L4, RTX 3060 |
| bfloat16 | ~12 GB | L4, A10G, RTX 4090 |

See the [main model card](https://huggingface.co/Akahsizrr/fuse-1-Lite) for full architecture details, training info, and technical report.