Mercury 3 tests large-scale text streaming

Inception launches Mercury 3, a diffusion language model reaching 1,068 tokens per second, featuring a 1-million-token context window and an API.

Ever since diffusion language models reached commercial use, one question has followed them: does their speed advantage survive as they grow? California startup Inception, cofounded by Stanford professor Stefano Ermon, answers with Mercury 3 and figures drawn from its own measurements. The model’s parameter count, which is not disclosed, is more than 40 times that of Mercury 1, and its output speed reaches 1,068 tokens per second on NVIDIA GPUs of an unspecified model. The context window grows to one million tokens, up from 260,000 for Mercury 2.5. A draft of the whole answer, refined in parallel

Autoregressive models write one token after another, each new token waiting on the previous one. A diffusion model starts from a rough draft of the entire answer, then corrects every position in parallel over a few passes. Fewer sequential steps mean faster output on the same GPUs, provided quality keeps up as the model grows. On that point, Inception says the score on its composite index tripled between Mercury 1 (mini) and Mercury 3, and that the scaling laws observed in autoregressive models also apply to diffusion. In the launch press release, Stefano Ermon places the limits of the approach in compute and data rather than in any fundamental barrier in the technique itself. An in-house index, renamed after the fact

The comparison with GPT-6 Luna and Claude Haiku 5.5, the speed-optimized small models from OpenAI and Anthropic, comes from tests run by Inception, with every model set to medium effort. The headline aggregate score, first presented as an “Intelligence Index,” was the subject of a clarification posted on 𝕏 after launch: it is a weighted average of twelve evaluations, including GDPval, SciCode, Tau3-Banking, IFBench, GPQA Diamond, Terminal-Bench v2.1 and BrowseComp+, and is distinct from Artificial Analysis’s index of the same name. The launch post now calls it “Model Quality,” and two tests on the list were renamed to avoid confusion with that organization’s own evaluations.

In the detailed table, as transcribed by MarkTechPost, Mercury 3 leads both rivals on eight of twelve tests. The widest gaps are in instruction following, at 76.84 on IFBench against 61.82 for GPT-6 Luna and 54.65 for Claude Haiku 5.5, and in agentic coding, at 51.08 on SWE-Bench Atlas against 33.33 and 32.26. Longer coding tasks and long-context retrieval still favor Claude Haiku 5.5, which comes out ahead on DeepSWE v1 (51.62 vs. 46.31) and MRCR v2 (78.40 vs. 75.50). By a thinner margin, GPT-6 Luna leads on AA-Omniscience-Public, which covers factual knowledge, and on AA-LCR v1.1, focused on long-context reasoning. The same table puts the rivals’ output speed at around 130 tokens per second for GPT-6 Luna and 240 for Claude Haiku 5.5, hence the claimed four- to eightfold gap.

Four early-access partners also weigh in on the post, with feedback published by Inception rather than independent evaluations. On 50 web navigation tasks drawn from Mind2Web, the TinyFish team says Mercury 3 completed half, significantly more than GPT-6.1-Sol by its account, without giving the latter’s score. For search on smart glasses, Octen AI measured a median generation time of 0.9 seconds and a full response in 2.2 seconds, while tutoring platform RevisionDojo says hints and graded answers come back in about a second. Cheaper output at launch, pricier at list

The official rate card sets Mercury 3 at $0.20 per million input tokens and $0.70 per million output tokens, but a 50% launch discount with no stated end date brings that to $0.10 and $0.35. Cached input tokens cost 90% less than the input price. For short prompts, GPT-6 Luna and Claude Haiku 5.5 are both billed $0.10 for input and $0.50 for output, with higher rates beyond 272,000 input tokens for the former and 100,000 prompt tokens for the latter. Token for token, Mercury 3 matches them on input and costs 30% less on output during the promotion, then twice as much on input and 40% more on output at list price.

Access is API-only, with no open weights, through Inception’s platform, OpenRouter, Baseten and Vercel’s AI Gateway. Compatible with the OpenAI API, the model answers to the `mercury-3` identifier and accepts a `reasoningeffort` parameter to set the depth of reasoning, according to the documentation. Every new account gets 100 million free tokens with no payment details required, and a chat interface serves as a no-code test bench. A deadline for Mercury 2, a setting to check for prompts

The arrival of Mercury 3 closes the door on earlier versions for newcomers: according to the models page, Mercury 2 and Mercury 2.5 are no longer offered to new users. Customers who used them before October 19, 2026 keep access until November 15, 2026, and enterprise clients remain covered by their existing agreements, with migration guidance available through their Inception representative or the documentation.

The platform’s terms of use also allow Inception to use submitted content, prompts included, to train its models. Opting out requires switching off the “Improve the model for everyone” option in the API Platform’s user settings, while enterprise offerings provide configurable data retention.