Diffusion LLMs Grow Up: Mercury 3 Scales 40x and Hits 1,068 Tokens per Second While Matching Frontier Speed Models
Synopsis
Whether diffusion language models can compete with autoregressive models on complex tasks had been an open question. Inception's Mercury 3 scales parameters more than 40x over Mercury 1 and runs at 1,068 tokens per second on NVIDIA GPUs; the company reports it matches or beats GPT-6 Luna and Claude Haiku 5.5 on several benchmarks while running 4-8x faster. In 50 live Mind2Web navigation tasks it completed 50%, well above GPT-6.1-Sol, using the same token volume and agent steps.
Interpretation
A diffusion language model is pushed to a scale where it can be compared directly with frontier speed-optimized models while keeping thousand-token-per-second generation. Mercury 1 showed diffusion LLMs could serve commercial applications, but whether they scale to complex use cases against autoregressive models remained open. Inception reports Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks and reaches 1,068 tokens per second on NVIDIA GPUs, 4-8x faster than those two.
The speed advantage is tested inside a real agent loop, not only as single-pass generation timing. Agents wait on the model at every step, so small delays compound quickly. In a production web-agent harness on 50 live Mind2Web navigation tasks, Mercury 3 completed 50%, significantly more than GPT-6.1-Sol, with the same token volume and agent steps and a fraction of the per-token cost.
The route is to keep scaling the diffusion model rather than switch to autoregressive decoding. Diffusion LLMs had largely been validated at small scale, with little evidence that scale and intelligence rise together. Total parameters grew more than 40x, and intelligence across standard benchmarks rose 3x from Mercury 1 (mini) to Mercury 3; Inception says the scaling laws behind autoregressive progress hold for diffusion too.
Perspective
The report covers Inception's Mercury 3 release and the benchmarks, production agent test and early-partner accounts it provides. It is aimed at researchers and engineers working on real-time inference and agent deployment, where speed, quality and context length combine. The stated next step is the API platform with 100 million free tokens and enterprise workload-fit evaluations.
Inception states that scaling laws hold for diffusion, a judgment resting on its own benchmarks and the reported 3x intelligence gain. The production agent test covers only 50 Mind2Web navigation tasks, with a 50% completion rate. Partner accounts are early tests and production observations without a shared evaluation protocol.