Editorial illustration for Inception Mercury 3: 1,068 tokens/sec on Nvidia GPUs, quality on par or better

Inception Mercury 3: 1,068 tokens/sec on Nvidia GPUs,...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

October 22, 2026 • 3 min read

Inception says its new Mercury 3 model can crank out more than 1,068 tokens per second on Nvidia GPUs, a speed gap wide enough that it claims a win over GPT-6 Luna and Claude Haiku 5.5 on eight of twelve benchmarks. The company calls Mercury 3 the most capable diffusion language model released so far, and it's betting that speed, not just raw accuracy, is what separates it from the pack.

The trick is architectural, not just a bigger GPU budget. GPT and Claude-style models write one word at a time, each token waiting on the last. Mercury 3 works more like the diffusion models behind image generation: it sketches a rough draft of the entire response, then refines it in passes, producing many tokens at once instead of queuing them up one by one. Google's technical report on DiffusionGemma lays out the same underlying concept.

Inception built Mercury 3 for the kind of apps where lag is the whole problem, autonomous agents, smart glasses, live customer support, where a model that thinks in word-by-word installments just can't keep pace. The model handles context windows up to 1 million tokens and plugs into existing pipelines through an OpenAI-compatible API.

According to Inception, Mercury 3 generates 1,068 tokens per second on Nvidia GPUs. In the company's comparison, Claude Haiku 5.5 manages about 240 and GPT-6 Luna about 130. No independent tests exist yet.

Inception's Mercury 3 diffusion model claims to beat GPT-6 Luna and Claude Haiku 5.5 at a fraction of the wait - THE DECODER

Why this matters

Inception's numbers come from Inception, tested against models the company picked and benchmarks it designed. That's worth remembering before anyone rewrites their inference stack around Mercury 3. Still, the architecture bet is real: diffusion models refine a whole draft at once instead of grinding out tokens one at a time, and if 1,068 tokens per second on Nvidia GPUs holds up under independent testing, that's a genuine shift for anyone building latency-sensitive products, chat interfaces, coding assistants, voice agents, anything where users feel every extra second.

For founders, the pitch is cost and speed without the usual quality trade-off, which is exactly the kind of claim that needs a third party to verify before it shapes product decisions. For researchers, the open question is whether diffusion-based generation scales as cleanly as autoregressive models have, or whether Mercury 3's gains shrink on harder reasoning tasks outside Inception's chosen comparisons. We'd watch for benchmarks run by labs with no stake in the outcome, and for real deployments where developers report their own throughput numbers instead of Inception's.