Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second

Inception has released Mercury 3, a diffusion language model. Instead of writing one token after another, it drafts the whole answer and refines all positions in parallel over a few passes. Inception reports 1,068 tokens per second on NVIDIA GPUs, a 1-million-token context window, and leads over OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 on 8 of 12 benchmarks in its own table.

In Inception's comparison, those rivals run at roughly 130 and 240 tokens per second. If those gaps hold up in practice, Mercury 3 changes what is possible for voice agents, live search and coding loops where waiting is the real cost.

The bigger claim is about scaling. Inception says Mercury 3 is more than 40 times larger than its first model, average quality rose about threefold, and speed stayed above 1,000 tokens per second. The open question for diffusion text models has always been whether they keep up as they grow.

We are genuinely interested in this one. Almost every major text model today writes sequentially, and that design sets a floor on how fast an answer can arrive. A credible second architecture would be healthy for the field, both for speed and for competition. Inception's biggest reported wins are in instruction following and agentic coding, including 51.08 on SWE-Bench Atlas against about 33 for both rivals. That is a large margin for a model sold on speed.

Now the skepticism the story requires. Every benchmark here was run by Inception at medium settings, the parameter count is not disclosed, there are no open weights, and the early customer stories are testimonials rather than independent evaluations. One partner says Mercury 3 beat GPT-6.1-Sol on web tasks but gives no number for Sol. We would not build a purchasing decision on that table alone.

The losses are telling, too. Haiku 5.5 wins DeepSWE v1 and the MRCR long-context retrieval test, and GPT-6 Luna edges ahead on two other knowledge and long-context measures. Long tasks and long-context retrieval look like the weak spots, and those are exactly the places where agents get into trouble. A fast model that loses the thread on the twentieth step can produce mistakes faster than humans can catch them.

That is the downside we would name if this scales. Speed is a safety property in reverse: when agents act at a thousand tokens a second, the window for a person to notice something is wrong shrinks. Cheap, fast generation also lowers the cost of spam and automated persuasion. None of that is unique to diffusion models, but faster engines make old problems arrive sooner.

Our read: Mercury 3 makes diffusion a real contender rather than a lab curiosity, and pricing at launch is aggressive. I'm watching for independent leaderboards to confirm the speed and coding numbers before calling it a shift.

Mercury 3 follows Mercury 1 in 2025, Mercury 2 and Mercury 2.5, which added tunable reasoning and a 260K context window. It is available through Inception's API, OpenRouter, Baseten and Vercel with an OpenAI-compatible interface. List pricing is $0.20 per million input tokens and $0.70 per million output, discounted 50% at launch. Inception also released Mercury Voice for voice agents and Mercury Decide for classification.