Inception Launches Mercury 3, a Diffusion LLM Claiming GPT-6 Luna Quality at 1,068 Tokens per Second

Inception, the Stanford spinout that has spent 2026 arguing that diffusion is a viable alternative to autoregressive language models, released Mercury 3 on October 19 and called it the most intelligent diffusion LLM ever released. The claim that matters is speed: the company reports 1,068 tokens per second on Nvidia GPUs, with a context window that has grown from 260,000 tokens in Mercury 2.5 to a full 1 million.

The pitch is architectural. Most large language models write one token at a time. Diffusion models draft many tokens in parallel and refine them over several passes, trading sequential dependency for throughput. Mercury 1, released in June 2025, was the first commercial proof that the approach could work; Mercury 2 added reasoning and Mercury 2.5 made that reasoning tunable. Inception says Mercury 3 is more than 40 times the size of Mercury 1 in total parameters, with benchmark intelligence up about 3 times across the same span, and co-founder Stefano Ermon argues the limits on scaling diffusion are compute and data rather than fundamental algorithmic barriers.

Pricing is aggressive. List rates are $0.20 per million input tokens and $0.70 per million output, halved to $0.10 and $0.35 for launch, with cached input at $0.01 during the promotion. The model is OpenAI API compatible, so switching does not require rewriting client code, and it is available through the Inception API platform, OpenRouter, Baseten and the Vercel AI Gateway. New accounts get 100 million free tokens. The constraints matter for buyers: the API is text-only, offers no embeddings endpoint and no seed parameter for reproducibility, and accepts temperature only between 0.5 and 1.0.

Whether it actually competes on quality is an open question. Inception says Mercury 3 outperforms the GPT-6 Luna model from OpenAI and Claude Haiku 5.5 from Anthropic across a range of benchmarks while running 4 to 8 times faster, but the announcement names no benchmark names, scores or methodology, and no independent evaluation has scored the model yet. Every quality figure in circulation is company-reported.

Early customers are at least running it in production-shaped settings. TinyFish AI says Mercury 3 completed 50 percent of 50 live Mind2Web browser navigation tasks in its web-agent harness, ahead of GPT-6.1-Sol with the same token budget and step count. Octen AI reports a median generation time of 0.9 seconds and a median end-to-end response of 2.2 seconds for smart-glasses search, and RevisionDojo says hints and marked answers now return in about a second. These are partner testimonials, not benchmarks.

The strategic read is about where latency pays. For a chat session, a thousand-token-per-second model is a novelty. For agents that make dozens of model calls per task - browser agents, real-time search, customer support - generation time compounds across steps, and that is exactly the workload Inception is targeting, together with the fact that speed-tier pricing usually meant a steep quality discount. If diffusion keeps scaling the way autoregressive models did, the fast lane stops being a compromise. Independent evals, not launch blog posts, will decide that. Inception says its next model is already in training.