Pinned post
🚀 Check out our Nemotron-Labs-Diffusion model!
We have been wondering about the true promise of diffusion LMs, especially when competing with strong AR models in terms of accuracy and with MTP methods in terms of efficiency.
💡 Nemotron-Labs-Diffusion is an important step toward answering this question, delivering a tri-mode LM that unifies AR, diffusion, and self-speculation (diffusion drafts, AR verifies) decoding within a single model.
🌟 Core Insights
🔸 AR and diffusion objectives can be mutually beneficial and harmonized within a single model.
🔸 Self-speculation, enabled by joint AR/diffusion training, can outperform Eagle3 in throughput.
🔸 Diffusion shows strong long-term potential for parallel decoding under an optimal sampler.
📊 Model performance: Our 8B/14B models match or exceed Qwen3-8B/14B accuracy while generating 6× more tokens per forward pass, leading to a 4× speedup in SGLang on GB200.
🤗 𝗛𝗙 𝗖𝗼𝗹𝗹𝗲𝗰𝘁𝗶𝗼𝗻 (3B/8B/14B base + instruct and 8B VLM):
lnkd.in/gej2VfN3
📰 𝗧𝗲𝗰𝗵 𝗥𝗲𝗽𝗼𝗿𝘁:
lnkd.in/gvS2Pwe8
We have been wondering about the true promise of diffusion LMs, especially when competing with strong AR models in terms of accuracy and with MTP methods in terms of efficiency.
💡 Nemotron-Labs-Diffusion is an important step toward answering this question, delivering a tri-mode LM that unifies AR, diffusion, and self-speculation (diffusion drafts, AR verifies) decoding within a single model.
🌟 Core Insights
🔸 AR and diffusion objectives can be mutually beneficial and harmonized within a single model.
🔸 Self-speculation, enabled by joint AR/diffusion training, can outperform Eagle3 in throughput.
🔸 Diffusion shows strong long-term potential for parallel decoding under an optimal sampler.
📊 Model performance: Our 8B/14B models match or exceed Qwen3-8B/14B accuracy while generating 6× more tokens per forward pass, leading to a 4× speedup in SGLang on GB200.
🤗 𝗛𝗙 𝗖𝗼𝗹𝗹𝗲𝗰𝘁𝗶𝗼𝗻 (3B/8B/14B base + instruct and 8B VLM):
lnkd.in/gej2VfN3
📰 𝗧𝗲𝗰𝗵 𝗥𝗲𝗽𝗼𝗿𝘁:
lnkd.in/gvS2Pwe8
0:38