By Hamza Faisal
Inception has released Mercury 3, its fourth diffusion language model and the most capable one ever built on the diffusion architecture — a model the company says outperforms OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 across a broad range of standard benchmarks, while generating text at more than 1,000 tokens per second. Independent write-ups report 1,068 tokens per second on Nvidia GPUs.
The headline claim is not the speed alone, nor the benchmark wins alone — it is that both arrived together. Mercury 3 is more than 40 times the size of Mercury 1, yet runs at comparable speeds, a combination that cuts against the long-standing assumption in AI that bigger models must be slower ones.
The trade-off that wasn't
Traditional large language models generate text sequentially, predicting one token at a time. Diffusion language models work differently: they generate and refine groups of tokens in parallel, like a photograph resolving out of noise. That parallelism is where the speed comes from, and it is why Inception can keep inference above 1,000 tokens per second even as the model grows.
The benchmarks come with the usual asterisk: they are self-reported, measured on Inception's own tables at medium settings. MarkTechPost's breakdown has Mercury 3 leading GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests — including SWE-Bench Atlas at 51.08, roughly 18 points ahead of both rivals — while trailing Haiku 5.5 on DeepSWE v1 by about 5 points. The company says Mercury 3 runs 4 to 8 times faster than those two frontier speed-optimized models while matching or beating their quality.
Diffusion, scaled
Since launching Mercury 1 in June 2025, Inception has shipped four models in roughly sixteen months. Mercury 2 added reasoning to the diffusion architecture; Mercury 2.5 introduced tunable reasoning and a 260,000-token context window. Mercury 3 is the largest leap in intelligence of any Mercury release, and it expands the context window to 1 million tokens — nearly quadrupling how much information the model can hold at once.
"Since we launched Mercury 1, the question we've heard most is whether diffusion can scale," said Stefano Ermon, Inception's co-founder. "We've already started training our next model, and we expect to continue pushing intelligence forward without sacrificing speed." The company is betting that the limits of diffusion models are compute and data, not fundamental algorithmic barriers.
Where it's already running
Mercury 3 is available through Inception's API — no open weights this time — and early customers are already deploying it in browser-use agents, interactive education, real-time search agents, and customer support, the exact applications where latency has historically forced developers to pick small models over smart ones.
That is the real story here. For developers building voice bots, agents, and search tools, speed-first models no longer have to mean capability-second. If diffusion can keep scaling without slowing down, the next frontier of model competition may not be about who builds the biggest model — but who builds the fastest smart one.
By Hamza Faisal
