Inception Labs says the biggest open question about diffusion language models now has an answer: they scale. These models draft a whole answer at once and refine it in passes, where conventional chatbots produce text one word after another. Inception’s new Mercury 3, announced October 19, is more than 40 times larger than Mercury 1 in total parameters, according to the company’s launch post.
Scaling was the doubt that kept this approach in the niche. Inception says benchmark intelligence rose 3x on the way up, and that the scaling laws behind conventional models hold for diffusion too. One detail in that figure: it runs from Mercury 1 (mini), not from Mercury 1 itself, though the surrounding text speaks of Mercury 1 throughout. The starting point is the small version.
The post is less clear on how good Mercury 3 is. Its opening says the model outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks. Its summary box says quality is “on par or better” than the same two models. Those are different claims. The post does not say which benchmarks fall into which group, and the weaker wording is the safer one to repeat.
The comparison set matters just as much. The post itself calls both rivals “frontier speed-optimized models,” meaning the fast tier of OpenAI’s and Anthropic’s lineups, not their best systems. Nothing in the post claims Mercury 3 beats a flagship model. It competes where cost and delay decide the purchase.
Speed is where the numbers are largest. Inception reports 1,068 tokens per second on NVIDIA GPUs, which it puts at 4x to 8x faster than those two models. The context window is 1M tokens, up from 260,000 in Mercury 2.5. At list price, a million tokens going in cost $0.20, and a million tokens coming out cost $0.70. Launch pricing is half that, $0.10 and $0.35, and the post gives no end date for the discount. Input served from cache costs 90% less than the standard input rate.
The one test against a larger model is a web-navigation run. The post says Mercury 3 completed 50% of 50 live Mind2Web tasks, “significantly more” than GPT-6.1-Sol, a different model from the two in the quality claim. It says Mercury 3 matched a frontier model on token volume and on the number of steps the agent took. The post says “we ran” the test inside a production web-agent harness, but the passage sits above a quote from Shuhao Zhang, cofounder of TinyFish AI, so who ran it is unclear. Either way, the sample is 50 tasks, the harness is not independent, and the post prints no score for GPT-6.1-Sol.
Early-access partners add support. Michael Tong, CTO of RevisionDojo, says hints and marked answers now come back in about a second. Octen AI’s Kuan Zou reports that, of the fast models Octen tested, Mercury 3 had the best balance. Its median generation time was 0.9 seconds, and the median full response took 2.2 seconds. These are customer testimonials hosted on the vendor’s blog. Every benchmark in the post is Inception’s own, and it links no independent evaluation.
The stake is the shape of the cheap, fast tier. Voice assistants, support bots, and agents that call a model at every step all lose time to latency, and that loss compounds across steps. If a second architecture can hold quality while running several times faster, buyers in that tier gain a real alternative to the two incumbents’ small models.
Mercury 3 is on Inception’s API platform and through OpenRouter, Baseten, and Vercel. It follows OpenAI’s API format, and Inception offers 100 million free tokens. A team running a latency-bound loop on a small model can test it in an afternoon, but should budget at the $0.20 and $0.70 list price, since the discount is temporary.
Inception Labs, “Diffusion at Scale: Introducing Mercury 3,” company blog, October 19, 2026.
