Inception introduces Mercury 3, a fast diffusion-based language model

Inception has launched Mercury 3, a diffusion-based large language model that the company says can generate more than 1,000 tokens per second while matching or exceeding several speed-focused models from OpenAI and Anthropic on selected benchmarks.

In an official blog post, Inception announces that Mercury 3 reaches 1,068 tokens per second on Nvidia GPUs. The company says the model is four to eight times faster than GPT-6 Luna, Claude Haiku 5.5, and GLM-5.3 Flash. Those comparisons and benchmark results have not been independently verified in the source material.

Mercury 3 is designed for applications where a delayed answer affects the user experience. Inception names voice and chat support, tutoring, search, data engineering, browser automation, and web agents as target use cases. The model supports a context window of up to one million tokens, which lets it process very long documents or larger collections of information in one request.

A different approach to generating text

Most widely used language models generate text sequentially. They predict one token, then use it to predict the next. This autoregressive approach can produce strong results, but it also creates a bottleneck because each new token depends on the previous one.

Diffusion models use a different process. They can work on multiple parts of a response in parallel and refine the output across several steps. This approach is already well established in image and video generation. Inception is applying it to text generation, with the goal of making better use of the parallel computing capacity of GPUs.

Jason Hiner reports for The Deep View that Inception has increased Mercury 3’s parameter count more than 40-fold compared with Mercury 1, released in 2025. The company does not disclose the model’s total number of parameters. Inception says performance across its standard benchmarks has tripled from Mercury 1 mini to Mercury 3.

According to The Deep View, Inception also used groups of Mercury 3 agents for recursive self-improvement work. The company says this process increased model speed by a further 16 percent. Details about the testing setup and the exact changes made to the model are not public.

Pricing and early production use

Mercury 3 costs $0.20 per million input tokens and $0.70 per million output tokens at its standard rate, according to Inception. An introductory price cuts those rates in half to $0.10 for input and $0.35 for output tokens. Cached input tokens receive a 90 percent discount from the listed input price.

Inception says Mercury 3 is available through its own API platform, as well as OpenRouter, Baseten, and Vercel. The service is compatible with the OpenAI API format, which could make migration easier for organizations already using software built around that standard.

The company cites early users including TinyFish AI, General Learning, Octen AI, and Lorikeet. TinyFish says Mercury 3 completed half of 50 live Mind2Web navigation tasks in its internal web-agent test, exceeding GPT-6.1-Sol under the same token and step limits. Octen says Mercury 3 achieved a median generation time of 0.9 seconds in its Sense API tests.

These customer accounts point to the main commercial question for diffusion language models. Their value may depend less on replacing the most capable frontier systems and more on enabling AI products that need responses quickly enough to feel conversational and responsive.

Sources

• Mercury 3: Diffusion at scale – Inception
• Inception launches Mercury 3, a diffusion-based LLM it claims is comparable to Luna and Haiku models but 4x to 8x faster, generating ~1,000 tokens per second – The Deep View