"We are all experts at something, but we’re all also beginners at something else."

— The Imitation Game (2014)

Arcana Qwen3 2.4B A0.6B

This is a MoE (Mixture of Experts) Qwen3 model which has total number of 2.4B parameters and 0.6B for each of 4 experts. All the expert models can be seen below. This model aims to provide more accurate results with more efficiency and less memory usage!

Expert Models:

Key Training Parameters (SFTConfig)

• per_device_train_batch_size = 2
• gradient_accumulation_steps = 4
• warmup_steps = 5
• num_train_epochs = 1
• learning_rate = 2e-5
• optim = "adamw_8bit"
• weight_decay = 0.01
• seed = 3407

Coding:

suayptalha/Qwen3-0.6B-Code-Expert

This model was fully fine-tuned with BF16 on first 20k rows of nvidia/OpenCodeReasoning dataset for 1 epoch.

Math:

suayptalha/Qwen3-0.6B-Math-Expert

This model was fully fine-tuned with BF16 on entire unsloth/OpenMathReasoning-mini dataset for 1 epoch.

Medical:

suayptalha/Qwen3-0.6B-Medical-Expert

This model was fully fine-tuned with BF16 on first 20k rows of FreedomIntelligence/medical-o1-reasoning-SFT dataset for 1 epoch.

Instruction Following:

Qwen/Qwen3-0.6B

Qwen/Qwen3-0.6B model was directly used for this expert, no fine-tune was applied.

Router Model:

The router model can be found here which was trained version of distilbert/distilbert-base-uncased on 7 different datasets.

Usage:

import torch
from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

local_dir = snapshot_download(
    repo_id="suayptalha/Qwen3-2.4B-A0.6B",
)

model = AutoModelForCausalLM.from_pretrained(
    local_dir,
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    local_dir,
)

model.to(device)
model.eval()

prompt = "I have pain in my chest, what should I do?"
messages = [{"role": "user", "content": prompt}]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

with torch.no_grad():
    output_ids = model.generate(
        text=prompt,
        max_new_tokens=1024,
        temperature=0.6,
        top_p=0.95,
    )
output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
print(output_text)

License:

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

Support: