Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution

Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion

Official implementation and model checkpoints for Orthrus, a dual-architecture framework that unifies the exact generation fidelity of autoregressive Large Language Models (LLMs) with the high-speed parallel token generation of diffusion models.

demo_orthrus.mp4

Model Zoo

All models use a Qwen3 backbone and guarantee strictly lossless generation.

Model	Base Model	HuggingFace	Avg. Speedup
Orthrus-Qwen3-1.7B	Qwen3-1.7B	🤗 HuggingFace	4.25×
Orthrus-Qwen3-4B	Qwen3-4.0B	🤗 HuggingFace	5.20×
Orthrus-Qwen3-8B	Qwen3-8.0B	🤗 HuggingFace	5.36×

Installation

uv pip install -e . uv pip install ninja packaging uv pip install flash-attn --no-build-isolation # or: pip install "flash-attn-4[cu13]" if your device supports it

We recommend uv for fast dependency resolution.

Quickstart

import torch from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer model = AutoModelForCausalLM.from_pretrained( "chiennv/Orthrus-Qwen3-8B", dtype=torch.bfloat16, device_map="cuda", attn_implementation="flash_attention_2", # use flash_attention_4 if your system does support trust_remote_code=True, ).eval() tokenizer = AutoTokenizer.from_pretrained("chiennv/Orthrus-Qwen3-8B") prompt = "Write a program to count the frequency of each word in a paragraph." 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=2048, use_diffusion_mode=True, streamer=TextStreamer(tokenizer, skip_prompt=True) # enable streaming generation )

Coming soon: Native integration with vLLM and SGLang is coming soon. Stay tuned!

Key Advantages

• Significant Inference Acceleration: Breaks the sequential bottleneck of standard autoregressive decoding, delivering up to a $7.8\times$ speedup on generation tasks.
• Strictly Lossless Generation: Employs an exact intra-model consensus mechanism to guarantee that the output matches the original base model's exact predictive distribution.
• Zero Redundant Memory Overhead: Both the autoregressive and diffusion views attend to the exact same high-fidelity Key-Value (KV) cache natively, resulting in only an $O(1)$ memory cache overhead.
• Parameter Efficient: Parallel generation capabilities are injected by fine-tuning only 16% of the total model parameters while keeping the base LLM strictly frozen.

Performance Comparison: Orthrus vs. Speculative Decoding

Orthrus outperforms speculative decoding methods like EAGLE-3, DFlash. By natively sharing the exact same KV cache across dual views, Orthrus avoids the redundant memory overhead of draft models, resulting in significantly higher token acceptance rates and faster inference times, especially as context length scales.

Left: Average verified tokens per forward pass compared to EAGLE-3 and DFlash. Right: Simulated generation time across scaling context lengths compared to DFlash.

Comparison with State-of-the-Art Diffusion Models

While recent diffusion language models (dLLMs) offer parallel decoding, they often suffer from significant conditional drift and severe accuracy degradation on complex reasoning tasks. Orthrus resolves this by decoupling parallel generation from sequential constraints, establishing a new state-of-the-art for parallel generation fidelity.

Throughput vs. Accuracy on MATH-500. Orthrus delivers a ~6x speedup over the Qwen3-8B baseline with strictly lossless performance, whereas adaptations like Fast-dLLM-v2 suffer significant accuracy drops.

Citation

If you find this model or architecture useful in your work, please cite our paper:

@misc{vannguyen2026orthrusmemoryefficientparalleltoken, title={Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion}, author={Chien Van Nguyen and Chaitra Hegde and Van Cuong Pham and Ryan A. Rossi and Franck Dernoncourt and Thien Huu Nguyen}, year={2026}, eprint={2605.12825}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2605.12825}, }