Diffusion LLM Mercury 3: 40 Times Larger Than Mercury 1, According to Inception – Parameter Count Kept Secret
The company behind the commercial diffusion models launched the fourth generation of Mercury on October 19, 2026, with a context window of 1 million tokens, a launch price of 10 cents per million input tokens, and benchmark wins against GPT-6 Luna and Claude Haiku 5.5 – but all performance figures come from Inception itself, and no independent evaluator has reproduced them.
The American company Inception launched Mercury 3 on Monday, October 19, 2026, describing the model as "the most capable diffusion LLM ever launched" – a characterization that appears in the company's own press release via Business Wire and is thus not an established assessment. What can be stated as fact, however, is that the model became available the same day via the Inception API, Baseten, OpenRouter, and Vercel, and that OpenRouter has confirmed that it went live the same day.
What Inception Claims
According to the press release, Mercury 3 runs at over 1,000 tokens per second on NVIDIA GPUs – MarkTechPost cites the figure of 1,068 tokens per second, as reported by Inception – beating GPT-6 Luna and Claude Haiku 5.5 while being 4–8 times faster. The model has a context window of 1 million tokens, up from 260,000 in its predecessor Mercury 2.5, so that far larger volumes of data can be processed in a single request.
Mercury 3 is the fourth release since Mercury 1 arrived in June 2025. Mercury 2 introduced reasoning in diffusion models, Mercury 2.5 added adjustable reasoning and a context window of 260,000 tokens, and Mercury 3 is now said to deliver the largest quality leap in the series – again, according to the company.
Co-founder Stefano Ermon announced the launch on X with the words that Mercury 3 is "the most intelligent diffusion LLM ever launched" and that it "beats GPT-6 Luna and Haiku 5.5 on quality while running at over 1,000 tokens per second." This too is the company's own characterization.
The Price
The list price is $0.20 per million input tokens and $0.70 per million output tokens. At launch the price is halved to $0.10/$0.35, and cached tokens are sold at a 90 percent discount on the input price, according to the press release.
The Benchmark Figures – and the Four Losses
On Inception's own table, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 on 8 out of 12 tests. According to the figures, the largest gains lie in instruction following and agentic coding: IFBench at 76.84, SWE-Bench Atlas at 51.08, and GPQA Diamond at 89.23.
But the table also shows defeats. Claude Haiku 5.5 wins DeepSWE v1 (51.62 versus 46.31) and MRCR v2 (78.40 versus 75.50), while GPT-6 Luna is narrowly ahead on AA-Omniscience-Public and AA-LCR v1.1. MarkTechPost points out that retrieval over long contexts and longer coding tasks are the model's weak points. All figures are reported by Inception at medium settings and come from the company's own table.
What Is Actually Verified
Precision matters here: all benchmark scores are supplied by Inception and have not been independently reproduced, as both AICoder and StartupFortune note. StartupFortune observes that the company's own comparisons have not been replicated by any independent evaluator, and that this pattern is not new.
Moreover, the speed claim rests on Inception's own measurements of the rivals – approximately 130 tokens per second for GPT-6 Luna and 240 for Haiku 5.5, according to AICoder – not on figures from OpenAI or Anthropic. None of the sources documents that these companies have confirmed or disputed the comparisons.
A further limitation: Inception has not published the parameter count of Mercury 3, only the claim that the model is over 40 times larger than Mercury 1. Nor has the exact NVIDIA GPU type been disclosed – Interesting Engineering writes only that the speed is said to have been achieved on "readily available" NVIDIA GPUs. The model is available only via API, with no open weights, so independent verification is limited to what Inception itself serves up.
The Scaling Claim
The question Inception wants to answer is whether diffusion architectures – which generate text via iterative refinement instead of left-to-right prediction – can be scaled up against autoregressive models without losing the speed advantage. "Since we launched Mercury 1, the question we have heard most often is whether diffusion can scale. Mercury 3 is over 40 times larger, substantially more capable, and still runs at over 1,000 tokens per second," Ermon said in the press release. "We are showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers."
He added that the company has already begun training the next model and expects to keep increasing intelligence without sacrificing speed. Note that this is based on the company's own testing; whether diffusion models can actually be scaled competitively further is a question none of the sources in this story can answer independently.
For now, what remains is a concrete offer – an affordable, fast model with 1 million tokens of context that is live on four platforms – alongside a list of claims that remain exactly that: claims.