Mercury 3 explained: the diffusion language model that writes 1,000 tokens a second

On 19 October 2026, the start-up Inception released Mercury 3, the newest and largest model in its Mercury family. It is a diffusion language model: instead of writing an answer one token at a time, like ChatGPT, Claude or Gemini, it produces a rough draft of the whole answer and refines every part of it at once. The result is speed. Inception reports 1,068 tokens per second, against roughly 130 for GPT-6 Luna and 240 for Claude Haiku 5.5 in its own comparison, and it now matches those small models on quality in most of the tests it published.

How a diffusion language model works

Every popular chatbot today is autoregressive: it predicts the next token, appends it, and repeats. Each token has to wait for the one before it, so a 1,000-token answer needs 1,000 steps in a row, however powerful the hardware. Diffusion models borrow the idea behind image generators such as Stable Diffusion: start from noise, then remove it over a few passes until a clean result appears.

Speed: the headline number

In practice, that turns a 500-token answer from a two- to four-second wait into half a second. For a person reading a chat window the difference is nice; for a voice assistant, a smart-glasses display or an agent that chains twenty model calls together, it changes what is possible.

Quality: close to the small frontier models

Inception compared Mercury 3 with the two most popular small models, at medium settings. It leads on 8 of 12 tests. The figures below are Inception's own; no independent leaderboard had published a score at the time of writing.

The pattern is consistent: Mercury 3 is strong at following instructions, science questions (89.2 on GPQA Diamond) and short coding tasks, and weaker at finding details in very long inputs and at longer, multi-step coding (Haiku 5.5 leads on DeepSWE, 51.6 to 46.3). For the full picture of where the small models sit, see the best AI models in October 2026 and Claude Haiku 5.5 explained.

Price and availability

Mercury 3 is available on Inception's API, OpenRouter, Baseten and Vercel AI Gateway. The API follows OpenAI's format, so most code only needs a new base URL and the model name mercury-3. There are no downloadable weights, so you cannot run it on your own servers. Mercury 2 and 2.5 are closed to new users, and existing customers keep access until 15 November 2026.

from openai import OpenAI

client = OpenAI(base_url="https://api.inceptionlabs.ai/v1", api_key="...")
reply = client.chat.completions.create(
    model="mercury-3",
    reasoning_effort="low",   # trade quality for even more speed
    messages=[{"role": "user", "content": "Summarise this ticket in one line: ..."}],
)
print(reply.choices[0].message.content)

Where speed actually matters

For analysis that needs deep reasoning over long documents, a larger autoregressive model is still the safer choice. Mercury 3 is best seen as a very fast worker inside a bigger system, not as the brain of it.

More on this month's models: the best AI models in October 2026, Claude Haiku 5.5 and Nvidia Nemotron 3.5 and Reflection Beam.

Sources: Inception's Mercury 3 announcement (19 October 2026), as reported by MarkTechPost and Unite.AI; OpenRouter's launch listing. All benchmark and speed figures are Inception's own and have not yet been independently reproduced.

Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 21 October 2026. About me