Inception's Mercury 3 is a diffusion LLM that claims 1,068 tokens a second and a 1M-token context

Inception, the company behind the Mercury family of diffusion language models, released Mercury 3 on 19 October and calls it “the most capable diffusion-based LLM released to date”. Diffusion LLMs refine a whole block of text in parallel passes instead of writing one token at a time, which is where the speed claims come from.

The numbers Inception gives

• Speed: 1,068 tokens per second on NVIDIA GPUs, which Inception says is 4 to 8 times faster than comparable speed-tier models.
• Context: 1M tokens, up from 260K for Mercury 2.5.
• Price: $0.20 per million input tokens and $0.70 per million output tokens. At launch it is 50% off, at $0.10 and $0.35, and cached tokens are 90% off the input price.
• Scale: “more than 40x larger than Mercury 1” in total parameters. Inception does not publish the parameter count.

Inception says Mercury 3 is “on par or better” than GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks. Those comparisons are Inception’s own runs, and there are no open weights to check them against.

Where you can use it

Mercury 3 is API-only. It is on the Inception API Platform, OpenRouter, Baseten and Vercel’s AI Gateway, and Inception says the API is OpenAI-compatible, so existing client code should only need a new base URL and model name. New API keys get 100 million free tokens.

The post quotes early-access partners in browser agents (TinyFish says Mercury 3 finished 50% of 50 live Mind2Web tasks, more than GPT-6.1 Sol in its harness), tutoring, search on smart glasses and customer support. These are partner testimonials, not independent tests.

Why it matters

Until now the question about diffusion LLMs was whether they could scale past small, fast, so-so models. Inception’s answer is yes: it says scores on standard benchmarks rose about 3x from Mercury 1 (mini) to Mercury 3. If that holds up outside Inception’s evals, latency-bound workloads, where agents wait on the model at every step, get a genuinely different option.