Inception has launched Mercury 3, a new "diffusion" language model it says runs at more than 1,000 tokens per second on widely available Nvidia GPUs. The Redwood City, Calif., startup announced it on Monday, Oct. 19, 2026.
Most chatbots write one token (a chunk of a word) at a time. Diffusion models draft many tokens in parallel and refine them, which is why they can be much faster.
What Inception says Mercury 3 does
- Quality: on par with or better than speed-focused models GPT-6 Luna and Claude Haiku 5.5 across standard benchmarks, while running 4 to 8 times faster. These are Inception's own results.
- Size: more than 40 times the size of Mercury 1 (June 2025), at comparable speed.
- Context: 1 million tokens, up from 260,000 for Mercury 2.5.
- Price: $0.20 per million input tokens and $0.70 per million output tokens, half off at launch ($0.10 and $0.35).
- Where: Inception API, OpenRouter, Baseten and Vercel; it is OpenAI API compatible.
"We're showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers," co-founder Stefano Ermon said. Inception says it has started training its next model. Its backers include Menlo Ventures, Microsoft's M12, Andrew Ng and Andrej Karpathy, Tesla's former AI director.
Why it matters: speed is the bottleneck for AI agents, which call a model again and again to finish one task. Related: websites blocking AI agents and Replit's switch to OpenAI.