Mercury 3: Diffusion language model writes an Asteroids game in 5 seconds

What this is about

On October 19, 2026, US startup Inception unveiled Mercury 3, which it calls its most capable diffusion language model (dLLM) to date. According to the company, it beats GPT-6 Luna and Claude Haiku 5.5 in 8 of 12 benchmarks while generating 1,068 tokens per second (tokens are the chunks of text that models work with). That would make it four to eight times faster than the comparison models. Mercury 3 is available in Inception Chat, through the Inception API, and on OpenRouter, Baseten and Vercel under the ID mercury-3.

How diffusion works for text

Inception is based in Redwood City, California. Its founders are researchers who worked on diffusion models. The CEO is Stanford professor Stefano Ermon, and the chief scientist is Vladimir Kuleshov. The company calls itself the creator of the first commercial dLLMs.

Most familiar language models are autoregressive: they predict the next token, append it, and start over. A 500-token answer takes 500 sequential passes, and token 300 has to wait for token 299. Diffusion models come from image generation (Stable Diffusion, Midjourney). For text, the model first produces a draft of the whole answer and "denoises" it in parallel over several passes. The number of passes is far smaller than the number of tokens, so the same hardware gets the job done faster.

The catches are that the answer length has to be set in advance and that the entire context must be reprocessed each time. That's why block diffusion is common: the answer is built block by block from left to right, diffusion runs inside each block, and finished blocks are cached, just as in autoregressive models. Inception doesn't say how Mercury 3 handles its 1-million-token context. One drawback: a model that doesn't write text linearly is harder to interpret, because its reasoning doesn't show up as the familiar chain. According to the source, that's an open question for all dLLMs.

What's new

Mercury 3 is Inception's fourth major release. The first Mercury appeared in 2025 as an experimental coding model. Mercury 2 added a reasoning mode and made TIME's list of the best inventions of the year in October 2026. Mercury 2.5 (September) introduced adjustable reasoning depth and a 260,000-token context window.

The core thesis is "diffusion scales": Mercury 3 is more than 40 times larger than Mercury 1 (mini), average quality across a benchmark suite rose roughly threefold, and speed stayed above 1,000 tokens per second. Inception doesn't disclose the parameter count. In an interview with SiliconANGLE, Ermon said the limits are compute and data, not fundamental algorithmic hurdles. Training of the successor is already underway. Kuleshov added on Hacker News that diffusion's advantage grows with the length of the reasoning: the more tokens a model generates before answering, the more it gains from parallelism.

Other details from the documentation: four reasoning levels, temperature from 0.5 to 1.0, tool calling and structured outputs. Input and output are text only, whereas GPT-6 Luna and Haiku 5.5 also process images.

Benchmarks – and what independent measurements show

Inception compares Mercury 3 with Claude Haiku 5.5 (released by Anthropic on October 7) and GPT-6 Luna, all at medium reasoning level. In the company's comparison, Mercury 3 leads on IFBench (76.84), GPQA Diamond (89.23) and SWE-Bench Atlas (51.08), among others, but trails Haiku 5.5 on DeepSWE v1 and MRCR v2. All the table figures come from Inception itself, and the comparison models deliver better but slower results at higher levels.

According to the source, the speed gap comes down to methodology: OpenRouter measures real user requests including network and queuing, while Inception measures raw throughput on its own GPUs. Artificial Analysis puts Haiku 5.5 at maximum settings at about 235 tokens per second and GPT-6 Luna at about 136. The source's conclusion: Mercury 3 looks strong for short agent steps, search and instruction following, but falls well behind on long multi-step terminal tasks.

Try it: Chat, playgrounds and Asteroids

Inception Chat (chat.inceptionlabs.ai) starts with Mercury 3 at medium level, with a web search toggle and a reasoning slider. The "Diffusion Effect" toggle shows how the answer emerges as a whole from random characters. Four playgrounds run without registration: Voice (a voice agent), Search (up to five parallel web searches), Decide (for example, ticket routing with urgency) and Code.

The Code Playground comes with a starter project containing index.html, style.css and game.js (a blue square you steer with the arrow keys). The Incrypted editorial team picked the "Asteroids" suggestion with the prompt "Make an Asteroids game: a ship that turns and thrusts, shoots, and rocks that split." The agent read the files, rewrote game.js, checked the code and edited index.html. Generation took up to 5 seconds and produced a game with score, lives, asteroid waves, blinking after a collision and restart with R. Controls: arrow keys or WASD, space bar to shoot. A second request (particle explosions, a shield power-up, a UFO every 30 seconds) took a few seconds.

The Live Design Editor also runs on Mercury 3: on a 1200×760 mockup of a café landing page, the model changes elements via text commands, such as a dark theme, a tea menu card starting at $2.50 and a dark green, rounded button. Changes show up as a preview; Enter applies them, Esc discards them.

Limitations

• The speed figure of 1,068 tokens/s applies to Inception's GPUs. In practice, it's about 243 via OpenRouter.
• According to the source, the model isn't yet suited for multi-hour terminal tasks and large refactorings.
• The comparison used the medium level. The competitors' higher levels are better but slower.
• Text only: no images or documents as images.
• No open weights and no details on parameters or architecture. You can't run the model locally.
• Mercury 2 and 2.5 will be shut down for existing customers on November 15, 2026.

Takeaway

According to the source, Mercury 3 is the first diffusion language model that, on average, matches the level of fast models from OpenAI and Anthropic. If the company's claims hold up, diffusion would be a real alternative wherever latency matters: browser and search agents, voice assistants, support and editor interfaces. The independent numbers do put the speed advantage in perspective, though, and the weak Terminal-Bench result shows its limits. For complex programming, flagships like Claude Opus 5.5 or GPT-6 Astra remain the better choice. Test your own prompts before switching in production.