Muse-3B

Muse-3B is a compact 3B chat language model from Muse Research Lab. It is built for helpful everyday conversation, writing, simple coding help, English/German/French assistance, and safe general-purpose responses.

Model Details

Model Developer: Muse Research Lab

Model Architecture: Muse-3B is an auto-regressive, Llama-style decoder-only transformer optimized for compact chat and general assistance.

Model	Params	Input modalities	Output modalities	Context Length	GQA	Shared Embeddings	Knowledge cutoff
Muse-3B	~3B	Multilingual text	Multilingual text and code	8,192 tokens	Yes	Yes	Not specified

Supported Languages: English, German, and French.

Status: This is an early compact chat model intended for lightweight assistant-style use and experimentation.

Capabilities

• General chat and question answering
• Writing, brainstorming, and rewriting
• Simple coding help and explanations
• Multilingual responses in English, German, and French
• Safe refusal behavior for harmful requests

Quickstart

pip install "transformers>=4.43.0" accelerate torch

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "Muse-research/Muse-3B"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are Muse-3B, a helpful chat assistant from Muse Research Lab."},
    {"role": "user", "content": "Hi, who are you?"},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
    )

response = tokenizer.decode(output_ids[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
print(response)

Intended Use

Muse-3B is intended for lightweight assistant-style use, including chat, drafting, summarization, simple programming support, and English/German/French/Italian/Spanish/Portuguese everyday help.

Limitations

• May produce incorrect or incomplete answers.
• May struggle with advanced reasoning, long coding tasks, or highly specialized domains.
• Multilingual support is useful but may be less reliable than English.
• Should not be used as the only source for medical, legal, financial, or safety-critical decisions.
• Applications should add their own safeguards when deployed to users.

Safety

Muse-3B is designed to be helpful while refusing clearly harmful requests. For production use, pair the model with application-level safety checks, monitoring, and domain-specific policies.

---

_{Built by Muse Research Lab}