Suzhou 3.1
A 9 billion parameter instruction-tuned language model by Tripplet Models. Suzhou 3.1 is an extremely strong AI Agent combining strong agent/tool-use capabilities with broad general knowledge.
Merge Details
- Method: SLERP (Spherical Linear Interpolation)
- Interpolation: Gradient blend across layers (attention vs MLP weighting)
Key Features
- 8.95B parameters — efficient enough to run on consumer hardware
- 262K context window
- Strong reasoning and chain-of-thought capabilities
- Tool calling and agent support (Hermes format)
- Multilingual support (29+ languages)
- Mixed attention architecture (linear + full attention layers)
Architecture
- Type: Causal Language Model
- Parameters: 8.95B
- Layers: 32
- Attention Heads: 16 (Q) / 4 (KV)
- Context Length: 262,144 tokens
- Precision: bfloat16
Quickstart
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "TrippletModels/Suzhou-3.1"
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt = "Explain the concept of attention in transformers."
messages = [
{"role": "system", "content": "You are Suzhou, created by Tripplet Models. You are a helpful assistant."},
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
**model_inputs,
max_new_tokens=512
)
generated_ids = [
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
Requirements
pip install transformers>=5.4.0 torch
License
Apache 2.0
Acknowledgments
Built on the work of:
- Tripplet Artificial Intelligence Research Institute (Tripplet)