Inception drops Mercury 3, a diffusion LLM hitting 1,000+ tokens per second

AI startup Inception just rolled out Mercury 3, a diffusion-based language model clocked at a blistering 1,068 output tokens per second on NVIDIA GPUs.

While standard autoregressive models write like a typewriter by guessing one token after another, diffusion models draft chunks of words simultaneously and polish them in parallel passes. Inception claims this architecture lets Mercury 3 outpace speed-oriented rivals like OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 by four to eight times.

Scaling the diffusion approach

Mercury 3 packs over 40 times the parameters of Mercury 1 and stretches its context window to 1 million tokens, up from 260,000 in version 2.5. Inception says overall intelligence jumped threefold across standard benchmarks without sacrificing its hallmark generation pace.

Co-founder Stefano Ermon noted that scaling text diffusion faces limits around compute and training data rather than fundamental algorithmic walls, adding that a successor model is already in the oven. Benchmark tables, however, show a nuanced picture: while Mercury 3 holds its own or wins on speed-optimized metrics, Luna and Haiku still edge it out on specific medium-effort reasoning evaluations.

On pricing, list rates sit at $0.20 per million input tokens and $0.70 per million output tokens, currently halved by a 50% launch discount down to $0.10 and $0.35, with prompt caching discounted by 90%.

Cutting latency in real workflows

Early partners are using the speed boost to tighten interactive feedback loops.

In browser navigation tests run by TinyFish AI across 50 live Mind2Web tasks, Mercury 3 finished 50% of the assignments, outperforming GPT-6.1-Sol using similar steps and token counts at lower cost.

Smart-glasses search startup Octen AI reported that model generation was previously the main bottleneck after an 80-millisecond retrieval. Mercury 3 trimmed median generation time to 0.9 seconds and median end-to-end response down to 2.2 seconds. Real-time tutoring platform RevisionDojo and support automation firm Lorikeet also plugged the model in to slash speech and text lag.

Why it matters

For builders running voice agents, conversational tutors, or multi-step web crawlers, Mercury 3 is well worth testing immediately. When an AI agent has to execute ten consecutive actions to finish a booking, waiting two seconds per step kills the experience. Dropping in a drop-in OpenAI-compatible endpoint that outputs over 1,000 tokens a second can make chained agent workflows feel instantaneous.

The broader race is shifting from raw benchmark supremacy to raw speed. OpenAI recently launched GPT-6 Astra Ultrafast, claiming up to 300 tokens per second, and secured 750 MW of dedicated capacity through Cerebras. While incumbents are trying to brute-force latency through dedicated hardware and specialized hosting tiers, Inception is trying to leapfrog them via architecture.

We think betting on diffusion is a sharp flanking maneuver, but teams should test carefully before swapping out production stacks. Extreme throughput does not guarantee immunity from compounding reasoning errors. Furthermore, as OpenAI noted in its own low-latency guidance, network connection overhead and external tool calls often dominate latency, meaning 1,000 tokens per second will only solve part of the end-to-end delay puzzle.