ai-forever/ZaryaOrthrus-1.7B

Zarya Orthrus 1.7B

This is a checkpoint of the Orthrus hybrid-architecture model based on Qwen3-1.7B. It was trained using the code available in the Zarya-Orthrus repository and the data available in the ZaryaOrthrusDataset-1.7B dataset.

Tokens Per Forward

The table below reports the speculative efficiency of the model, measured as the number of tokens generated per autoregressive forward pass.

Model	Params	code	code-ru	creative-en	gec-en	gec-ru	math	math-ru	poetry-en	poetry-ru	qa	qa-ru	wmt ru-en
ai-forever/ZaryaOrthrus-0.6B	0.8B	3.39±0.16	4.29±0.14	1.95±0.12	4.22±0.16	3.87±0.20	4.62±0.15	3.95±0.14	1.77±0.07	2.56±0.21	1.96±0.07	1.70±0.24	2.00±0.10
ai-forever/ZaryaOrthrus-1.7B	2.1B	3.47±0.14	3.97±0.13	2.02±0.10	4.26±0.16	3.25±0.17	4.96±0.16	4.29±0.14	1.84±0.04	2.29±0.14	2.27±0.06	2.09±0.20	2.45±0.12
chiennv/Orthrus-Qwen3-1.7B	2.1B	3.26±0.25	3.57±0.23	1.95±0.15	4.96±0.29	2.43±0.15	8.13±0.68	4.12±0.28	1.60±0.04	1.96±0.21	2.00±0.07	1.58±0.14	2.08±0.09

Getting Started

The following code demonstrates how to generate a response using the model:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
 
model_id = "ai-forever/ZaryaOrthrus-1.7B"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16, device_map="cuda",
    attn_implementation="flash_attention_2",
    trust_remote_code=True,
).eval()
tokenizer = AutoTokenizer.from_pretrained(model_id)
 
prompt = "Explain me what is Special Theory of Relativity, who invented it and what is its essence."
messages = [{"role": "system", "content": ""}, {"role": "user", "content": prompt}]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True, enable_thinking=False).input_ids
output_ids = model.generate(
    input_ids=input_ids.to(model.device), 
    max_new_tokens=512,
    use_diffusion_mode=True, 
    streamer=TextStreamer(tokenizer, skip_prompt=True) # enable streaming generation
)