New AI text diffusion models break speed barriers by pulling words from noise

On Thursday, Inception Labs released Mercury Coder, a new AI language model that uses diffusion techniques to generate text faster than conventional models. Unlike traditional models that create text word by word—such as the kind that powers ChatGPT—diffusion-based models like Mercury produce entire responses simultaneously, refining them from an initially masked state into coherent text.

Traditional large language models build text from left to right, one token at a time. They use a technique called "autoregression." Each word must wait for all previous words before appearing. Inspired by techniques from image-generation models like Stable Diffusion, DALL-E, and Midjourney, text diffusion language models like LLaDA (developed by researchers from Renmin University and Ant Group) and Mercury use a masking-based approach. These models begin with fully obscured content and gradually "denoise" the output, revealing all parts of the response at once.

While image diffusion models add continuous noise to pixel values, text diffusion models can't apply continuous noise to discrete tokens (chunks of text data). Instead, they replace tokens with special mask tokens as the text equivalent of noise. In LLaDA, the masking probability controls the noise level, with high masking representing high noise and low masking representing low noise. The diffusion process moves from high noise to low noise. Though LLaDA describes this using masking terminology and Mercury uses noise terminology, both apply a similar concept to text generation rooted in diffusion.

A chart provided by Inception Labs showing self-reported scores on the Artificial Analysis coding index. Credit: Inception Labs

Much like the creation of an image synthesis model, researchers build text diffusion models by training a neural network on partially obscured data, having the model predict the most likely completion and then comparing the results with the actual answer. If the model gets it correct, connections in the neural net that led to the correct answer get reinforced. After enough examples, the model can generate outputs with high enough plausibility to be useful for tasks like coding (although, so far, they still confabulate frequently on many topics).

According to Inception Labs, its approach allows the model to refine outputs and address mistakes because it isn't limited to considering only previously generated text. This parallel processing enables Mercury's reported 1,000-plus tokens per second generation speed on Nvidia H100 GPUs.

These diffusion models maintain performance faster than or comparable to similarly sized conventional models. LLaDA's researchers report their 8 billion parameter model performs similarly to LLaMA3 8B across various benchmarks, with competitive results on tasks like MMLU, ARC, and GSM8K.

However, Mercury claims dramatic speed improvements. Their Mercury Coder Mini scores 88.0 percent on HumanEval and 77.1 percent on MBPP—comparable to GPT-4o Mini—while reportedly operating at 1,109 tokens per second compared to GPT-4o Mini's 59 tokens per second. This represents roughly a 19x speed advantage over GPT-4o Mini while maintaining similar performance on coding benchmarks.

Mercury's documentation states its models run "at over 1,000 tokens/sec on Nvidia H100s, a speed previously possible only using custom chips" from specialized hardware providers like Groq, Cerebras, and SambaNova. When compared to other speed-optimized models, the claimed advantage remains significant—Mercury Coder Mini is reportedly about 5.5x faster than Gemini 2.0 Flash-Lite (201 tokens/second) and 18x faster than Claude 3.5 Haiku (61 tokens/second).

Opening a potential new frontier in LLMs

Diffusion models do involve some trade-offs. They typically need multiple forward passes through the network to generate a complete response, unlike traditional models that need just one pass per token. However, because diffusion models process all tokens in parallel, they achieve higher throughput despite this overhead.

Inception thinks the speed advantages could impact code completion tools where instant response may affect developer productivity, conversational AI applications, resource-limited environments like mobile applications, and AI agents that need to respond quickly.

If diffusion-based language models maintain quality while improving speed, they might change how AI text generation develops. So far, AI researchers have been open to new approaches.

Independent AI researcher Simon Willison told Ars Technica, "I love that people are experimenting with alternative architectures to transformers, it's yet another illustration of how much of the space of LLMs we haven't even started to explore yet."

On X, former OpenAI researcher Andrej Karpathy wrote about Inception, "This model has the potential to be different, and possibly showcase new, unique psychology, or new strengths and weaknesses. I encourage people to try it out!"

Questions remain about whether larger diffusion models can match the performance of models like GPT-4o and Claude 3.7 Sonnet, produce reliable results without many confabulations, and if the approach can handle increasingly complex simulated reasoning tasks. For now, these models may offer an alternative for smaller AI language models that doesn't seem to sacrifice capability for speed.

You can try Mercury Coder yourself on Inception's demo site, and you can download code for LLaDA or try a demo on Hugging Face.

Opening a potential new frontier in LLMs

nproxy.org