News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Mercury

Mercury

Diffusion-based language models for fast code generation

Inception Labs released Mercury in June 2025 as a family of large language models built on diffusion principles. Designed for coding tasks, these models predict multiple tokens in parallel to achieve high throughput speeds while maintaining quality comparable to leading alternatives.

Share on WhatsApp

  • Uses diffusion to predict multiple tokens in parallel
  • Mini variant reaches 1109 tokens/sec on H100 GPUs
  • Ranks second in quality on Copilot Arena
  • Outperforms speed-optimized models by up to 10x
  • Released by Inception Labs in June 2025

What it is

Mercury is a new generation of commercial-scale large language models developed by Inception Labs. Released in June 2025, the initial lineup focuses on coding applications under the name Mercury Coder. The models use a diffusion approach rather than standard autoregressive methods, allowing them to generate text much faster. Two sizes are available initially: Mini and Small. These models aim to balance speed and quality, targeting developers who need rapid code completion and generation capabilities.

How it came about

Inception Labs introduced Mercury in a technical report published in June 2025. The team included researchers such as Samar Khanna, Siddhant Kharbanda, and Aditya Grover. This release marked their first set of diffusion-based large language models designed specifically for coding tasks. The launch included the release of a public API and a free playground for users to test the models. The development focused on creating a new speed-quality frontier for language models used in software engineering.

Inside the model

The Mercury models are parameterized using the Transformer architecture but are trained to predict multiple tokens in parallel. This diffusion-based design differs from traditional models that generate one token at a time. By predicting several tokens simultaneously, the system achieves significantly higher throughput. The architecture allows for ultra-fast inference speeds while maintaining the structural benefits of Transformer networks. This parallel prediction mechanism is central to its performance claims.

What it can do

Independent evaluations by Artificial Analysis show that Mercury Coder Mini reaches 1109 tokens per second on NVIDIA H100 GPUs. The Small variant achieves 737 tokens per second on the same hardware. These speeds outperform other speed-optimized frontier models by up to ten times on average. Despite the focus on speed, the models maintain comparable quality to leading alternatives. On Copilot Arena, a real-world validation platform for developers, Mercury ranks second in quality and is the fastest model overall across various code benchmarks and languages.

How to use it

Inception Labs provides access to Mercury through a public API. Users can also interact with the models via a free playground hosted by the company. The initial release includes the Mini and Small variants of Mercury Coder. The models are described as commercial-scale, suggesting they are intended for professional and enterprise use cases in addition to individual developer testing.

Questions people ask about Mercury

What makes Mercury different from other language models?

Mercury uses a diffusion-based approach that predicts multiple tokens in parallel, unlike traditional models that generate text one token at a time. This design allows for much faster inference speeds while maintaining high quality.

How fast is the Mercury Coder Mini model?

According to independent evaluations, Mercury Coder Mini achieves a throughput of 1109 tokens per second when running on NVIDIA H100 GPUs, making it one of the fastest models available for coding tasks.

Who developed the Mercury models?

Mercury was developed by Inception Labs, a research organization. The team included authors such as Aditya Grover, Stefano Ermon, and Volodymyr Kuleshov, among others.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Mercury yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2506.17298. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.