Diffusion model claims 1,107 tok/s in Mercury 2.5

Inception announced Mercury 2.5, its new diffusion language model, on September 8 with two headline figures: 1,107 tokens per second on widely available NVIDIA GPUs, and an intelligence score 40% higher than Mercury 2’s.

In its launch post, also summarized on Threads, Inception says the model performs on par with Gemini 3.5 Flash and GPT 5.6 Luna Low and runs much faster than both. The company’s press release lists a slightly different set of peers, calling Mercury 2.5 “comparable to cost-optimized frontier models like GPT-5.6 Luna, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5”. Both comparisons come from Inception itself, and no outside benchmark has backed them so far.

List pricing is $0.20 / $0.75 per million tokens. At launch Inception is charging 80% less, $0.04 / $0.15. The context window goes from 128K to 260K tokens, and the model adds tunable reasoning, native tool use, JSON mode and parallel tool calls.

The speed comes from how diffusion models generate text. An autoregressive LLM writes one token after another. Mercury starts with a rough, noisy draft that covers many positions at once and refines all of them over a series of denoising steps until the output settles. Inception’s cost argument follows from this. With autoregression, cost and latency grow with reasoning depth, so a model that thinks longer is also slower and more expensive. A diffusion model produces many tokens per step, and the company says it does this on standard NVIDIA GPUs without specialized silicon.

The first Mercury shipped in February 2025, billed as the first commercially available diffusion LLM. Inception said Mercury 2 matched Claude Haiku and GPT Mini on intelligence at roughly 10x their throughput. Since Mercury 2 launched, the company says, usage has grown to thousands of developers and dozens of enterprises, mainly in search, voice and coding agents.

Inception’s technical report from June 2025 listed Mercury Coder Mini at 1,109 tokens per second on NVIDIA H100s, based on independent evaluations by Artificial Analysis.