Qwen3.6-25B-A3B

This is a compressed checkpoint derived from Qwen/Qwen3.6-35B-A3B.

Overview

• Base model: Qwen/Qwen3.6-35B-A3B
• Total parameter count: 34.66B -> 24.97B
• Layers: 40
• Active experts per token: 8
• Format: standard Transformers safetensor shards with tokenizer, generation config, and chat template included

The repo is set up for direct from_pretrained(...) loading.

Quick start

Use a recent Transformers release that supports Qwen3.6 MoE. This checkpoint was produced and validated with transformers==5.5.4.

pip install -U torch transformers==5.5.4 accelerate torchvision pillow

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Jaso1024/Qwen3.6-25B-A3B"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",
)

messages = [{"role": "user", "content": "Solve: If 3 notebooks cost $12, how much do 8 cost? End with ####."}]
try:
    prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, enable_thinking=False)
except TypeError:
    prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Notes

• The uploaded config.json already reflects the compressed architecture.
• This is still a large bf16 checkpoint, so practical inference typically needs high-memory GPU hardware or multi-device offload.
• License and usage terms should be treated as inherited from the upstream base model.