tripplet-research/taipei-3.1

Taipei 3.1

By Tripplet AI (Tripplet Artificial General Intelligence Research Institute)

Taipei 3.1 is a massive 754B parameter Mixture-of-Experts language model built on the GLM-5.1 architecture, delivering frontier-level performance in both English and Chinese.

Model Details

• Parameters: 753.9B (MoE)
• Architecture: GLM MoE with Dense-Sparse Attention (glm_moe_dsa)
• Languages: English, Chinese
• License: MIT
• Precision: bfloat16
• Model Size: ~1.51 TB

Key Features

• Frontier-class MoE architecture with dense-sparse attention
• Bilingual (English + Chinese) with strong multilingual capabilities
• Optimized for instruction following and conversational tasks
• Competitive with leading closed-source models

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("tripplet-research/taipei-3.1", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "tripplet-research/taipei-3.1",
    trust_remote_code=True,
    torch_dtype="bfloat16",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Hello, tell me about yourself."}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
output = model.generate(inputs.to(model.device), max_new_tokens=512)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Hardware Requirements

This model requires significant compute resources:

• Minimum: 8x A100 80GB or equivalent (~640GB+ VRAM)
• Recommended: 16x A100 80GB or H100 cluster for full precision inference

About Tripplet AI

Tripplet Artificial General Intelligence Research Institute is dedicated to advancing the frontiers of artificial general intelligence through open research and model development.

License

MIT