Mercury 3, Inception's diffusion LLM, now in FastMetal chat
Mercury 3, which Inception released on October 19, 2026, is now available in FastMetal's chat. It is a diffusion language model: instead of writing one token after another, it drafts the whole response and refines it in parallel. Inception calls it the most capable diffusion LLM released so far.
What's new
- Speed. Inception measured 1,068 tokens per second on NVIDIA GPUs and puts it at 4 to 8 times faster than comparable models. That matters most where waiting compounds: agent loops, and anything a person is watching.
- Bigger, with a longer context. More than 40 times the size of the original Mercury, and the context grows from 260K tokens on Mercury 2.5 to 1M.
- Quality. Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks, without publishing the individual scores.
- Tools and structured answers. We confirmed tool calling and schema-constrained JSON answers on real requests before launch.
- Text only. It does not read images or PDFs.
It answers Japanese questions in Japanese as well as English; we checked before launch.
Pricing
$0.11 per 1M input tokens and $0.385 per 1M output tokens. Repeated context, such as conversation history, is cached and billed at $0.011 per 1M. Tokens the model spends reasoning bill as output. Prices can change, so check the model page for current rates. Billing is prepaid, so you never spend more than you have loaded; see pricing.
How to try it
- Create a FastMetal account.
- Add credit.
- Open chat, pick Mercury 3 from the model list, and ask away.
Compare real answers first
The model page shows Mercury 3's actual answers to our 43 shared prompts: summarization, bug fixing, translation and more. Set them beside Mercury 2.5 to see what changed.
Browse the full model catalog.