Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second
MarkTechPost Jean-marc Mommessin
Inception’s Mercury 3 can answer at 1,068 tokens a second and handle 1M-token prompts. It’s API-only and beats rivals on 8 of 12 tests, but not all.
Based on reporting by MarkTechPost, Jean-marc Mommessin — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Inception has shipped Mercury 3, the latest version of its Mercury diffusion language model, and the headline is speed. The company says it can generate at 1,068 tokens per second on NVIDIA GPUs while taking in prompts as long as 1 million tokens. That puts it squarely in the “move fast first” camp, which is exactly where a lot of agent and voice work lives.
Mercury 3 doesn’t write like a normal autoregressive model. Instead of producing text one token at a time, it starts with a rough draft and refines the whole answer in parallel. That matters because fewer sequential steps usually means less waiting. Inception’s pitch is that this approach still scales, even as the model gets larger.
The company says Mercury 3 has more than 40 times the parameters of Mercury 1, which launched in 2025, and that average benchmark quality has risen about 3x across those early generations. It is also the fourth Mercury release, after Mercury 1, Mercury 2 and Mercury 2.5. The last of those added tunable reasoning and a 260K context window, according to Inception.
On Inception’s own benchmark table, Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests, all run in medium settings. Its strongest showing is SWE-Bench Atlas, where it scores 51.08, roughly 18 points ahead of both rivals. IFBench is another bright spot at 76.84, and GPQA Diamond lands at 89.23.
But the misses are real. Haiku 5.5 does better on DeepSWE v1 and MRCR v2, while GPT-6 Luna edges Mercury 3 on AA-Omniscience-Public and AA-LCR v1.1. So this is not a clean sweep; it is a fast model with clear strengths in instruction following and agentic coding, and softer edges in long-context retrieval and longer coding tasks. Mercury 3 is available through the Inception API Platform, OpenRouter, Baseten and Vercel, with API access only and no open weights or quantized builds.
My take — AI-written commentary, not fact-checked reporting
The market keeps pretending benchmark tables are the whole story, then acts surprised when the fastest model is still a tradeoff machine. API-only diffusion models are neat, but they’re also a reminder that the industry loves selling speed before it solves the boring parts like reliability and access. Benchmarks can flatter almost anything if you look long enough.
Read more about this at: MarkTechPost