What Is Inception Mercury 2? The Diffusion LLM Explained

Mercury 2 is a reasoning language model from Inception that generates text by diffusion. It launched on February 24, 2026, and Inception called it the fastest reasoning LLM, at 1,009 tokens per second on NVIDIA Blackwell GPUs. Eight months later the lineup has moved on. Mercury 2.5 arrived in September, and Mercury 3 on October 19, 2026. Mercury 2 is closed to new users, and existing customers lose access on November 15. Below: how diffusion decoding works and where Mercury 2 users should move.

What is Inception Mercury 2, in one sentence?

Mercury 2 is a diffusion LLM (dLLM) that drafts many tokens at once and refines them over a few steps, so it runs several times faster than token-by-token models.

Inception's launch post by CEO Stefano Ermon puts it as "less typewriter, more editor revising a full draft at once." The company built its first commercial dLLMs before Mercury 2. Its 2025 technical report describes Mercury Coder, a Transformer trained to predict multiple tokens in parallel. Mercury 2 added tunable reasoning to that recipe.

How does Mercury 2 compare with the newer Mercury models?

All three are Inception diffusion models. Each newer one has a larger context window, and Mercury 3 is the cheapest per token.

Sources: Inception's launch posts for Mercury 2 and Mercury 2.5. Mercury 3 data comes from its launch post and the models page.

How does a diffusion LLM generate text?

A standard LLM is autoregressive. It writes the first token, then the next one conditioned on the first, until the answer is done. Each token waits for the one before it, so long answers take time no matter how fast the chip is.

A diffusion LLM starts from a rough draft of the whole response and refines it. Inception's launch post says Mercury 2 produces "multiple tokens simultaneously" and converges over a small number of steps. Each step updates many positions at once, which suits GPUs well. Image models such as Stable Diffusion work on the same principle, applied to pixels.

The speed gain also changes the reasoning tradeoff. Reasoning models normally pay for quality with longer chains of thought, which means more latency. Inception argues that diffusion lets a model reason inside a real-time budget. That is a vendor argument, and the benchmark charts behind it are Inception's own.

What could Mercury 2 do, and what did it cost?

The launch post listed four features: tunable reasoning, a 128K context window, native tool use and schema-aligned JSON output. The API is compatible with the OpenAI API, so existing clients only need a new base URL and key. Inception says the models also work through LiteLLM and LangChain.

Mercury 2 launched at $0.25 per million input tokens and $0.75 per million output tokens. Inception said its quality was "competitive with leading speed-optimized models." Customers quoted in the post include Zed for code suggestions and Skyvern, which said it was at least twice as fast as GPT-5.2. Wispr Flow cited transcript cleanup, and OpenCall cited voice agents. Later in the year the Mercury 2.5 post listed two hosts beyond Inception's own API: Baseten and OpenRouter.

Where do diffusion models fit in an agent workflow?

Agents call the model many times per task, so latency per call adds up. Inception pitches Mercury for the calls around a main model. Its examples range from autocomplete and search pipelines to voice turns and subagents.

The Mercury 2.5 post gives one concrete case. Augment Code uses Mercury inside its coding agent for context compaction and two other side tasks, routing and tool search. Inception reports that moving compaction to Mercury cut its latency by 82%, from about 150 seconds to 27. Cost fell by 90%. These are Inception's numbers about a customer, and Augment's own write-up was not part of our sources.

We would copy this pattern: keep a frontier model for planning and hard edits, and hand the frequent, well-defined calls to a fast model. Summarizing a long session to fit the context window is a good first candidate.

What should you read in the fine print?

Inception is the source for every speed figure and quality comparison here, and for the customer quotes. The only independent numbers predate Mercury 2. The 2025 technical report cites independent speed measurements by Artificial Analysis for Mercury Coder: 1,109 tokens per second for Mini and 737 for Small, on H100 GPUs. We found no equally detailed independent test of Mercury 2.

Mercury 2 also has a short shelf life. The models page says Mercury 2 and 2.5 are no longer available to new users. Existing customers with usage before October 19, 2026 keep access until November 15, 2026. Enterprise contracts are handled separately. Inception points everyone else to Mercury 3. It says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 on its chosen benchmarks while running 4 to 8 times faster.

The bottom line

If you use Mercury 2 today, plan the move to mercury-3 before November 15, 2026. The API is the same shape, and Mercury 3 is cheaper per token with a much larger window. If you have never tried a diffusion model, start with a high-volume side task such as summarization or routing. Measure it against your current fast model before you trust the vendor numbers.

How do you keep agent memory when the fast model changes?

Mercury 2 is being retired less than nine months after launch, so the model under an agent can change fast. Augment's compaction case exposes a second limit. A compaction summary helps one long session fit in the window. It does not carry project facts into tomorrow's session, and it disappears with the session. What an agent learns about your codebase needs a store that outlives both the session and the model.

memU is one such store for coding agents. A scheduled background task slices new session logs into jobs. The agent distills those jobs into memory and skill Markdown, which you can read and edit. memU itself makes no LLM calls. It stores and embeds what the agent prepared, then retrieves it on request. A standing instruction in CLAUDE.md or AGENTS.md tells the agent to retrieve before answering. memU sits in the host, outside the Mercury API, so its memory stays put when you swap the model behind Claude Code or Codex.

The hosted version is free and syncs across devices. Get an API key at memu.so and send your agent one message:

Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is YOUR_MEMU_API_KEY.

The self-hosted path and the host support matrix are in the memU GitHub repository.

FAQ

Is Mercury 2 still available?

Only for existing customers, and only until November 15, 2026. Inception's models page says Mercury 2 and Mercury 2.5 are closed to new users. Mercury 3 replaces them.

How fast is Mercury 2?

Inception reported 1,009 tokens per second on NVIDIA Blackwell GPUs at launch. That is the company's own measurement.

How much did Mercury 2 cost?

Mercury 2 launched at $0.25 per million input tokens and $0.75 per million output tokens. Mercury 3 lists at $0.20 and $0.70, with a 50% launch discount.

Is Mercury 2 compatible with the OpenAI API?

Yes. Inception says its Mercury models are OpenAI API compatible. Existing clients need a new base URL and key, plus the new model name.

What is the difference between a diffusion LLM and a normal LLM?

A normal LLM writes one token after another. A diffusion LLM drafts a whole block and refines many tokens in parallel over a few steps, which is where its speed comes from.