holgt/qwen3-0.6b-burn

Qwen3-0.6B for Burn

This repository contains Qwen3-0.6B weights in formats compatible with the Burn deep learning framework.

Model Details

• Base Model: Qwen/Qwen3-0.6B
• Parameters: 0.6B
• Architecture: Qwen3 (decoder-only transformer)
• License: Apache-2.0

Configuration

Parameter	Value
hidden_size	1024
num_hidden_layers	28
num_attention_heads	16
num_key_value_heads	8
intermediate_size	3072
vocab_size	151936
max_position_embeddings	40960
rope_theta	1000000
rms_norm_eps	1e-6

Available Formats

File	Format	Size	Description
model.safetensors	HuggingFace SafeTensors	1.4 GB	Original BF16 weights from Qwen
model.bpk	Burn Burnpack	1.4 GB	Converted for Burn (BF16)
tokenizer.json	HuggingFace Tokenizers	11 MB	Tokenizer file

Usage with qwen3-burn

Using .bpk format (recommended)

use qwen3_burn::{Qwen3Config, Qwen3ForCausalLM, Qwen3Tokenizer};
use burn::backend::candle::{Candle, CandleDevice};
use half::bf16;

type Backend = Candle<bf16, i64>;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let device = CandleDevice::metal(0); // or CandleDevice::Cpu

    // Load tokenizer
    let tokenizer = Qwen3Tokenizer::from_file("tokenizer.json")?;

    // Initialize model with 0.6B config preset
    let model = Qwen3Config::qwen3_0_6b()
        .init_causal_lm::<Backend>(&device)
        .with_weights("model.bpk")?;  // or "model.safetensors"

    // Generate text
    let (input_ids, _) = tokenizer.encode("Hello, world!")?;
    let input_tensor = Tensor::from_data(&input_ids, &device).unsqueeze();

    let output = model.generate_with_cache(
        input_tensor,
        50,    // max_new_tokens
        0.0,   // temperature (0 = greedy, >0 = sampling)
        0.9,   // top_p
        50,    // top_k
    );

    let text = tokenizer.decode(&output.to_data().to_vec())?;
    println!("{}", text);

    Ok(())
}

Performance

On Apple M-series with Metal backend:

• ~25 tokens/sec with greedy decoding (temperature=0)
• Model loading: ~2-3 seconds

Acknowledgments

• Original model by Alibaba Qwen Team
• Burn framework by Tracel AI