Inception Releases Mercury 3: a Diffusion LLM 8x Faster Than GPT-6 Luna

Diffusion language models have long been seen as fast but weak: fine for autocomplete, questionable for serious tasks. Mercury 3, released on October 19, tries to shed that stigma. According to Jiqizhixin, the model is more than 40 times larger than the previous generation, generates text roughly 8 times faster than GPT-6 Luna and outperforms it on a number of benchmarks. The output price is $0.35 per million tokens.

The project is backed by a team led by Stanford University professor Stefano Ermon, one of the authors of research on diffusion models. Back in 2025, Inception Labs showed Mercury Coder and claimed more than a thousand tokens per second on an H100. At the time, it looked like an impressive speed demo for a narrow task. Now it is an attempt to move into the territory of general-purpose models.

Why a 40x increase in size matters

The central question about diffusion LLMs has always been the same: will the speed advantage hold once the model gets big? Autoregressive models generate a response token by token, and each step waits for the previous one. A diffusion scheme starts from a noised or masked sequence and refines the whole thing over several passes, processing many positions in parallel. The fewer passes it takes to reach a decent result, the higher the throughput.

Scaling a model up by more than 40 times usually hurts latency. More parameters mean more computation on every pass. If Mercury 3 is still 8 times faster than GPT-6 Luna at that size, parallel generation really does offset the cost of scale. For the industry, this is an important signal: the architectural advantage has not dissolved as the model has matured.

Economics: $0.35 as an argument

A price of $0.35 per million output tokens hits developers where it hurts. Almost every provider charges more for output tokens than for input tokens, and agentic workflows, code generation and long reasoning chains burn through thousands of them per request. A fast model frees up the GPU sooner, and the provider can pass that on as a lower price.

Key reported parameters:

• Release: October 19, 2026
• Type: diffusion language model (Mercury 3)
• Scale increase: more than 40 times over the previous generation
• Speed: roughly 8 times higher than GPT-6 Luna
• Price: $0.35 per million output tokens
• Quality: outperforms GPT-6 Luna on benchmarks

It is worth looking separately at which tests Mercury 3 actually won. Speed on coding and structured-output tasks and quality on long reasoning chains are measured differently, and a general headline like "beat it on benchmarks" does not yet mean a win everywhere.

Who should test it first

The first candidates for a trial are obvious. They are teams for whom response latency directly affects the product: IDE assistants with autocomplete, voice agents where a one-second pause breaks the conversation, and bulk-generation pipelines that count tokens in the billions per month.

The practical plan is simple. Take your real traffic, run it through Mercury 3 and your current model, and compare three metrics: time to full response, cost per thousand requests, and the share of responses that pass your automated checks. If quality holds up on your task, the savings on output tokens will pay for the integration very quickly. If not, you will at least have a clear picture of where diffusion models are already good enough and where they are not yet.

For the market as a whole, Mercury 3 poses an uncomfortable question to the leaders: if parallel generation scales, then optimizing autoregressive models through speculative decoding and caching runs into a ceiling, while the rival has no such ceiling.