Mercury 3 is a diffusion language model that Inception says produces 1,068 tokens per second on Nvidia GPUs, beats OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 on eight of 12 benchmarks, and carries a one-million-token context window. Nearly every number there comes from the company selling the model. That is how launches work, and it is why the announcement reads as a marketing document first and a technical report second.
The mechanism is the interesting part. Diffusion is the technique that made Stable Diffusion possible for images: instead of painting pixel by pixel, the model starts from noise and refines the whole picture across several passes. Mercury 3 applies the same idea to text, drafting a rough version of the whole answer and refining many tokens in parallel instead of emitting one at a time. Here Inception is on firm ground: parallel refinement really does remove the sequential bottleneck that makes autoregressive decoding slow.
What Inception Is Actually Selling
Strip away the framing and the pitch is a latency product. Inception is not arguing that Mercury 3 is the smartest model available. It is arguing that the model is fast enough that waiting stops being a design constraint. The company's own summary describes the quality as on par or better, which is a notably softer claim than the benchmark headline suggests.
The numbers Inception published:
- 1,068 tokens per second, against roughly 240 for Claude Haiku 5.5 and about 130 for GPT-6 Luna in the company's own comparison.
- Eight wins out of 12 internal benchmarks, with all three models evaluated at medium settings.
- 76.84 on IFBench, about 15 points ahead of GPT-6 Luna on instruction following.
- 51.08 on SWE-Bench Atlas, against 33.33 and 32.26 for the two rivals.
- More than 40 times the parameters of Mercury 1, with no actual parameter count disclosed.
Early-access partners share speed results, not intelligence results, which tells you where the value sits. Octen AI, which builds search for smart glasses, says its search step takes about 80 milliseconds and that the language model was the bottleneck. Founder Kuan Zou puts the stakes bluntly: on smart glasses, an answer that takes five seconds feels broken. With Mercury 3, Octen reports a median response of 0.9 seconds and a 2.2-second end-to-end flow. Education company General Learning says it now returns hints and corrections in about a second.
The agent evidence is thinner. TinyFish AI gave Mercury 3 fifty navigation tasks on real websites. It solved 25. Inception says that beat GPT-6.1-Sol, without publishing the competitor's score. Fifty tasks is a demo, not an evaluation.
The Benchmark Table Is Doing a Lot of Work
Leading eight of 12 tests is a real result if it holds up. It does not hold everywhere. Claude Haiku 5.5 wins the longer coding benchmark DeepSWE and the long-text retrieval test. GPT-6 Luna takes two tests from Artificial Analysis. The honest summary: Inception built a model unusually strong at instruction following and autonomous coding, but behind on the capability enterprises lean on hardest, pulling the right fact out of a long context.
Two details deserve scrutiny. No independent benchmarks existed at launch; every comparison traces back to Inception's own harness. And the claim of being 40 times larger while running just as fast depends on a parameter count the company declines to publish. Scaling claims without disclosed scale are marketing, not measurement.
Google works the same lane and is more willing to admit the trade-off: its open DiffusionGemma runs up to four times faster than the standard Gemma 4 while delivering slightly weaker results. When a lab with Google's resources frames diffusion as a speed-for-quality bargain, a startup claiming no quality penalty should expect hard questions.
The Price Cut Is the Tell
Mercury 3 lists at 0.20 dollars per million input tokens and 0.70 dollars per million output tokens. At launch Inception is halving both, to 0.10 and 0.35. The company has not said when the promotion ends, which is worth noticing, because at standard pricing Mercury 3 costs more per output token than both rivals on shorter inputs, where GPT-6 Luna and Claude Haiku 5.5 each charge 0.50 dollars.
Read that alongside the speed claim and the strategy becomes clear. Inception is selling throughput and subsidising it to buy developer mindshare in agent and voice workloads, where a pipeline makes hundreds of calls and a 200-millisecond saving per call compounds into something users feel. For very long prompts the calculus flips in Mercury 3's favour, since the million-token window avoids the surcharges rivals apply to large inputs. That advantage rewards one specific workload shape, not a general one.
Three questions to answer before moving a production pipeline:
- Does the speed hold on your hardware, your prompt lengths and your production load, or only on the vendor's benchmark rig?
- Does quality hold on your long-context retrieval tasks, the one area where the published table shows a loss?
- What happens to unit economics when the promotional pricing ends?
None of this makes Mercury 3 a bad model, and the underlying bet is credible. Diffusion text generation has real momentum, Inception's founders come from the work that popularised diffusion, and the company shipped Mercury 2 in February as what it called the first diffusion model to reason through intermediate steps, backed by the venture arms of Nvidia, Microsoft and Snowflake.
What deserves pushback is the presentation. A speed record measured in the vendor's own harness, an undisclosed parameter count behind a 40 times scaling claim, a temporary 50 percent discount presented next to permanent list prices, and a headline that omits the two tests where the cheaper rival wins. Speed is not an outcome. It is one input into an outcome, and the distance between 1,068 tokens per second on a launch chart and a responsive product people trust is where all the actual engineering still lives.
Comments