AI startup Inception has unveiled Mercury 3, a diffusion language model that challenges traditional autoregressive approaches by refining many tokens simultaneously, achieving faster speeds and strong benchmark performance against GPT-6 Luna and others.
- Mercury 3 uses diffusion to generate text faster than autoregressive models
- Model supports a one-million-token context window, vastly larger than before
- Launch pricing cuts highlight speed and efficiency for AI development
What happened
Inception, a Redwood City-based AI lab, released its latest language model, Mercury 3, which employs a diffusion-based architecture distinct from the autoregressive models dominating the market. Unlike conventional models that generate text token by token, Mercury 3 begins with noisy output and refines multiple tokens concurrently until coherent text emerges.
This new approach allows Mercury 3 to produce over 1,000 tokens per second on Nvidia hardware, reportedly running four to eight times faster than competitors including OpenAI’s GPT-6 Luna, Anthropic’s Claude Haiku 5.5, and Zhipu’s GLM-5.3 Flash. The model features a dramatically expanded context window of up to one million tokens and is accessible through APIs compatible with OpenAI’s, facilitating straightforward adoption by developers.
Why it matters
The diffusion technique demonstrated by Mercury 3 represents a notable divergence from the traditional autoregressive paradigm, potentially reducing latency and improving scalability for AI-powered applications. Faster token generation speeds enable more efficient deployment in environments requiring rapid and repeated model calls such as coding assistants and interactive agents.
Additionally, Mercury 3's expanded context window allows it to manage significantly larger bodies of text in a single pass, broadening the scope of use cases it can handle effectively. Inception's pricing strategy, with initial discounts lowering costs below some competitors, further underscores their focus on speed and operational efficiency as key differentiators in a competitive market.
What to watch next
Independent benchmarking will be crucial to validating the performance claims made by Inception, as the current speed and capability metrics are based solely on the company’s own tests. Industry observers will look for third-party evaluations comparing Mercury 3’s diffusion approach directly against autoregressive models under diverse workloads.
Furthermore, adoption and integration within AI ecosystems through platforms like OpenRouter, Baseten, and Vercel’s AI Gateway will indicate developer acceptance and practical utility. The longevity of Inception’s promotional pricing and the impact of recursive self-improvement techniques on the model's future iterations will also be key factors shaping the market reception and evolution of diffusion language models.