Share This Article
Mercury AI model could be a breakthrough in the market
Startup Inception Labs has released the Mercury AI model. The developer claims that it processes information 10 times faster than GPT-4o.
The most common architecture variant of existing language models is LLM. At the same time, developers are regularly looking for alternatives that can compete with existing products. Inception Labs has introduced a new approach based on diffusion. Mercury is the first solution of its kind to focus on the needs of commercial use.
Testing on the Artificial Analysis platform has shown excellent results. Mercury is ten times faster than current models on the market. The performance of the solution is more than a thousand tokens per second. NVIDIA H100 GPUs were used for the test. Previously, such results could only be achieved on specialised equipment.
Features of the model
Transformers play a key role in the creation of LLM text. They generate tokens sequentially. Diffusion models work differently:
- AI, in this case, creates text simultaneously;
- the generation process is carried out from deep analysis to detail;
- the initial stage is noise processing, which later becomes cleaner;
- as a result, a single stream of tokens is formed.
Diffusion is most commonly used for videos and images. This approach is at work in the Sora and Midjourney AI. However, the Mercury model has moved away from this principle, opening up new possibilities for the use of LLM. One of the main advantages of diffuse models is accelerated content generation.
In standard coding studies, Inception Labs’ Mercury outperformed the results of popular AI models. These are GPT-4o Mini, the next-generation Gemini and Claude 3.5 Haiku.
The developers claim that the diffuse approach to AI has several advantages, especially in logical processes. The Mercury model structures answers without the constraints of previous tokens. This improves the quality of text generation. In addition, Mercury can refine the input data on a regular basis. Such a feature reduces the likelihood of errors in the generation of the result.
According to Inception Labs, their development can compete with models on the market. Modern models for processing complex data require a large amount of computing resources. This is due to the creation of long chains of logic, which results in significant costs. There is also a delay in the output of the results.
The AI model used by Mercury is of higher quality. The company plans to work on improving the solution while a test version is on the market. This will allow users to evaluate the full capabilities of the model. In the future, Mercury will be available for commercial use with enhanced functionality.