Skip to content
Neural Daily

ModelsNews

Mercury 3 hits 1,068 tok/s as a diffusion LLM

On October 19 Inception released Mercury 3, its most capable diffusion language model so far. The company reports 1,068 tokens per second on NVIDIA GPUs and a 1 million token context. The OpenAI-compatible API is also on OpenRouter, Baseten and Vercel.

By: Mikola Máté Patrik · · 2 min read

Inception cover image: Diffusion at scale, Introducing Mercury 3, white text over a wave pattern on black
Image: Inception · Press kit · Official blog share image · Source

What changed versus Mercury 2.5

Mercury 3 is more than 40x larger than Mercury 1. Inception says intelligence on standard benchmarks rose about 3x from Mercury 1 mini to Mercury 3, while speed stayed high. Context grew from 260K to 1M tokens (Inception).

List price is $0.20 / $0.70 per million input / output tokens. At launch it is 50% off at $0.10 / $0.35. Cached input is 90% off. Inception claims parity with or better than GPT-6 Luna and Claude Haiku 5.5 at 4-8x the speed.

Inception benchmark table: Mercury 3 compared with GPT-6 Luna and Claude Haiku 5.5 on speed, price, capability, long context, coding and search
Inception · Press kit · Official benchmark table from the Mercury 3 blog post · Source

Source: Inception. The company's own comparison table for Mercury 3 Medium, GPT-6 Luna Medium and Claude Haiku 5.5 Medium.

Where partners already run it

Early-access partners use it for voice and chat support, real-time tutoring, data engineering and browser agents. In Inception's own web-agent harness on 50 live Mind2Web tasks, Mercury 3 completed 50%, more than GPT-6.1 Sol at similar step counts and token volume.

Quoted feedback from TinyFish, RevisionDojo, Octen AI and Lorikeet highlights lower latency without giving up quality. In Octen's Sense API tests, median generation was 0.9 seconds and median end-to-end response 2.2 seconds.

What a diffusion LLM does differently

Autoregressive models write one token at a time. Diffusion models fill masked positions in parallel over several rounds, which is where the speed comes from, especially when reasoning models produce long answers. Mercury 3 weights are not open. The API is OpenAI-compatible, with 100 million free tokens to try.

Practical note

All figures are vendor-reported. For latency-sensitive agents, voice apps or browser automation, run a head-to-head A/B against your current autoregressive model on the same harness.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles