Inception has launched Mercury 3, a diffusion-based language model that generates text up to eight times faster than traditional AI models. Co-founder Stefano Ermon announced the release on October 19.
Unlike conventional language models that generate text one word at a time, Mercury 3 uses a diffusion approach. It refines multiple text blocks in parallel, similar to an editor revising an entire draft at once. This method allows it to produce over 1,000 tokens per second on widely available NVIDIA GPUs.
Inception claims Mercury 3 outperforms both GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks. However, the company has not published specific benchmark names or scores yet. As a result, these claims remain unverified by independent testing.
1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec. https://t.co/TYevDl43Lm
— Stefano Ermon (@StefanoErmon) October 19, 2026
The model also features a 1 million token context window, a major jump from 260,000 tokens in Mercury 2.5. Additionally, Inception says Mercury 3 is over 40 times larger than the original Mercury model while maintaining comparable processing speeds.
Mercury 3 targets applications where delays hurt usability. These include interactive tutoring, live translation, writing assistants, and automated software agents. Early-access customer RevisionDojo reported that Mercury 3 cut student feedback response times to roughly one second.
The model is now available through the Inception API, Baseten, OpenRouter, and Vercel. It also supports the OpenAI API format, so developers can integrate it without major code changes.
Inception has moved quickly since launching Mercury 1 in June 2025. Mercury 2 added reasoning capabilities, and Mercury 2.5 introduced adjustable reasoning with a 260,000 token context window. The company raised $50 million last year from investors including Microsoft, NVIDIA, and Snowflake. Ermon also confirmed that Inception has already started training its next model.
Still, the diffusion approach faces questions about long-term competitiveness. Google’s Gemini Diffusion showed similar promise but has received no public updates since its initial presentation. Whether diffusion-based language models can keep pace with rapidly improving autoregressive rivals remains an open question.