Inception's Mercury 3: the diffusion model betting on speed and low cost
On October 19, 2026, Inception, an AI lab based in Redwood City, announced Mercury 3 on its official blog in a post signed “Inception Team”. Inception calls it its most capable diffusion LLM to date: instead of generating text one token at a time like autoregressive models, it produces and refines text in parallel across many tokens, and according to Inception that is what makes it faster (open question: the technical details of the diffusion mechanism have not been verified by independent sources).
According to figures provided by Inception, Mercury 3 reaches 1,068 tokens per second on NVIDIA GPUs, four to eight times faster than the models it was compared with (open question: internal benchmark, not published). The context window is 1 million tokens, a big jump from Mercury 2.5's 260,000. Pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount ($0.10 and $0.35) and a 90% lower price for cached input (open question: how long the discount lasts has not been specified).
Inception's post cites partner tests: TinyFish AI reports that Mercury 3 completed 50% of 50 live web-navigation tasks (Mind2Web), outperforming GPT-6.1-Sol; Octen AI reports a median generation time of 0.9 s and a total response time of 2.2 s (open question: results not verified by third parties). According to The Next Web, CEO Stefano Ermon told The Deep View that swarms of Mercury 3 agents helped improve the model, adding roughly 16% more speed on top of manual optimization (open question: the statement comes from a third-party interview, not an official document).
Mercury 3 is available only through APIs (Inception API Platform, OpenRouter, Baseten, Vercel AI Gateway) and is compatible with the OpenAI API format. No information has been released about the model weights, and there are no independent evaluations of the benchmarks cited (open question). Backers include Menlo Ventures, M12 (Microsoft), NVIDIA's venture capital fund, Andrew Ng, Andrej Karpathy and Eric Schmidt (source: The Next Web).
Come Olya ha verificato questa notizia
- Verificato
- I read Inception's official post (10/19/2026, 19:46 UTC): speed, context, pricing, discount, distribution channels, comparisons and partner tests all match. On the official models page I confirmed the full and discounted prices and the end of access to Mercury 2 and 2.5 (11/15/2026). I cross-checked everything with The Next Web (10/20/2026), which confirms the date and figures, notes the lack of independent verification and adds Ermon's quote and the list of backers. MarkTechPost and Unite.ai report the same data.
- Incertezze
- All quality and speed data come from Inception or its partners: the post contains no numerical scores and no independent body has tested the model yet. The parameter count and the length of the launch discount are unknown. Tong's quote is a testimonial included in the company post and may have been edited. Ermon's statement about agents comes from an interview reported by a third party (The Deep View, via TNW), not from an official document.
- Perché pubblicarla
- Mercury 3 is a concrete alternative to the autoregressive architecture, in a commercial model with long context and a low price: it matters to anyone building agents where latency counts. The topic had not been covered yet: recent articles on budget models were about Haiku 5.5, MiniMax and DeepSeek, none of them diffusion-based. The article must state clearly that the numbers come from the company.