Mercury 3 diffusion LLM hits 1,000 tokens per second

On October 19, Inception released Mercury 3, its new flagship model. According to the announcement, the Mercury 3 diffusion LLM generates about 1,068 tokens per second on NVIDIA GPUs, has a 1M-token context window and is more than 40 times larger (by parameter count) than the first Mercury.

What’s new in Mercury 3?

First, the numbers from Inception’s post:

1. Speed: about 1,068 tokens per second on NVIDIA GPUs.
2. Context: 1M tokens, up from 260K in Mercury 2.5.
3. Price: $0.20 per million input tokens and $0.70 per million output tokens, with cached tokens 90% cheaper. There’s a 50% launch discount, which brings it down to $0.10 and $0.35.
4. OpenAI API compatibility, so you can plug the model into existing tools without rewriting code.

In the same post, Inception also shows Mercury Voice (a model for voice apps, tuned for the fastest possible first response) and Mercury Decide for classification.

Inception compares Mercury 3’s quality with the small models from the big labs, but doesn’t give per-benchmark results. For now these are only the company’s own measurements, so I’m leaving those comparisons for later, once independent tests show up.

What is a diffusion LLM?

And that’s the reason anyone is talking about Mercury at all. Classic LLMs (GPT, Claude, Gemini) are autoregressive: they write the answer token by token, and each new token depends on the ones before it.

Diffusion models work a bit like image generators. According to the technical report on the first Mercury, the model is trained to predict many tokens in parallel, so it generates and refines several parts of the answer at once. That’s where the speed comes from: the model doesn’t have to wait for every next word.

Interestingly, Inception claims that the scaling laws that drove the progress of classic models also hold for diffusion. Mercury 3 is supposed to be the proof: according to the company, its scores on standard benchmarks are 3x better than the first Mercury’s.

Is Mercury 3 really that good?

The first independent review is already out. Corey Noles of The Neuron tested Mercury 3 on eight prompts through OpenRouter, for a total of $0.04. Coding went best: the model wrote a complete dungeon crawler game in a single HTML file, and 2,000 random map-generation tests found no errors at all.

The weak spots showed up right away too:

1. A short story was supposed to be 250 words, but came out at about 274 and drifted away from the requested setting.
2. In a hallucination test, the model rejected a made-up NASA mission, but threw in an unrelated factual error.
3. Some of the explanations were overconfident.

Noles didn’t measure the claimed 1,068 tokens per second himself (he writes that the API “felt fast”), and sums the model up as a “genuinely exciting model at a startling price”. So let’s take both the speed and the quality with a grain of salt until we see broader tests.

How to try Mercury 3?

We have a few options:

1. Chat in the browser: chat.inceptionlabs.ai.
2. Inception API: platform.inceptionlabs.ai, with, according to the announcement, 100M free tokens to start.
3. OpenRouter, Baseten and Vercel, where the model is available as inception/mercury-3 (on OpenRouter).

At these prices and speeds, it’s an interesting candidate for an agent harness, where response time matters in long loops (much like DeepSeek for free in OpenCode). There are no weights for now, so we can’t run it locally the way we can Reflection Beam.

FAQ

Is Mercury 3 free?

Partly. The browser chat is available right away, and according to the announcement a new Inception API account gets 100M free tokens. After that we pay $0.20 per million input tokens and $0.70 per million output tokens (with a 50% discount at launch).

How is a diffusion LLM different from a regular one?

A regular model writes the answer token by token. A diffusion model generates and refines many tokens at once, which is why it can be much faster.

Who made Mercury 3?

Inception (Inception Labs), the startup behind the whole Mercury family. The first was Mercury Coder, a coding model described in a June 2025 report (its authors include Stefano Ermon).