Shuttle-3 - AI Language Models Tool

Overview

Shuttle-3 is a fine-tuned, 72–73B-parameter causal language model released by ShuttleAI in October 2024. It was built from the Qwen-2.5-72B-Instruct base and optimized for multi-turn chat, multilingual instruction following, role-playing scenarios, and reasoning-heavy tasks. The project explicitly aims to emulate the high-quality prose and conversational alignment found in Claude 3–style models while retaining strong code and multilingual capabilities. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3?utm_source=openai)) Shuttle-3 is offered on Hugging Face (safetensors/BF16 formats) and is the backing model for ShuttleAI’s ShuttleChat product (beta), which demonstrates the model in interactive chatbot/character applications. The model README documents its ChatML prompting format and summarizes the fine-tuning setup used to produce the instruction-following behavior. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3?utm_source=openai))

Model Statistics

• Downloads: 173
• Likes: 39
• Pipeline: text-generation
• Parameters: 72.7B

License: other

Model Details

Architecture and base: Shuttle-3 is a causal LLM fine-tuned from Qwen/Qwen2.5-72B-Instruct and reported at roughly 72.7B parameters (commonly rounded to 72–73B). It exposes a text-generation pipeline and is distributed as safetensors in BF16 precision. ([promptlayer.com](https://www.promptlayer.com/models/shuttle-3?utm_source=openai)) Fine-tuning & training: According to the model README, Shuttle-3 was post-trained on ~130 million tokens for approximately 12 hours using 4× A100 PCIe GPUs, targeting improved role-playing, multi-turn consistency, and Claude-style prose. Prompting uses a ChatML-style format (system/user/assistant tokens) to structure multi-role dialogues. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3/blob/main/README.md?utm_source=openai)) Capabilities and intended uses: Shuttle-3 is positioned for complex chat, multilingual instruction-following, coding assistance, role-play characters, and agent-like workflows (tool integration patterns are demonstrated by the ShuttleChat product). The model is offered from the ShuttleAI organization on Hugging Face and is intended both for local/hosted inference and integration via ShuttleAI’s platform. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3?utm_source=openai))

Key Features

• Fine-tuned from Qwen-2.5-72B-Instruct for instruction-following and chat.
• Trained on extensive role-playing data to improve multi-turn persona consistency.
• Designed to emulate Claude 3–style prose and conversational alignment.
• Multilingual and code-aware pretraining for translation and coding assistance.
• Uses ChatML prompting (system/user/assistant markers) for structured dialogs.
• Distributed as BF16 safetensors for efficient inference on modern accelerators.

Example Usage

Example (python):

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

# Replace with a device config appropriate for your hardware (this example uses device_map='auto')
model_id = "shuttleai/shuttle-3"

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

gen = pipeline("text-generation", model=model, tokenizer=tokenizer)

prompt = (
    "<|im_start|>system\nYou are a helpful assistant.\n<|im_end|>"
    "<|im_start|>user\nWrite a friendly two-paragraph description of a travel app.\n<|im_end|>"
)

result = gen(prompt, max_new_tokens=250, do_sample=False)
print(result[0]["generated_text"])

Benchmarks

Parameter count: ≈72–73B parameters (Source: https://huggingface.co/shuttleai/shuttle-3)

Fine-tuning data: 130 million tokens (post-training) (Source: https://huggingface.co/shuttleai/shuttle-3/blob/main/README.md)

Training hardware & time: Trained ~12 hours on 4× A100 PCIe GPUs (Source: https://huggingface.co/shuttleai/shuttle-3/blob/main/README.md)

Downloads (last month, Hugging Face): 309 downloads (last month, as reported on Hugging Face) (Source: https://huggingface.co/shuttleai/shuttle-3)

Tensor / file format: BF16, safetensors (recommended for inference) (Source: https://huggingface.co/shuttleai/shuttle-3)

Key Information

• Category: Language Models
• Type: AI Language Models Tool