← Back to Blog
New model

Mercury 2.5 diffusion LLM added to FastMetal (since retired)

FastMetal

FastMetal no longer offers Mercury 2.5. The model was withdrawn upstream in November 2026, and it is no longer in the catalog. This post records what it was when we added it. See the model catalog for what is available today.

Inception's Mercury 2.5 (mercury-2.5) joined FastMetal on September 8, 2026. It is a diffusion LLM (dLLM): instead of writing tokens one at a time from the start, it drafts several tokens in parallel and refines them. Inception pointed to that design as the reason it called Mercury its fastest reasoning model.

What it offered

  • Parallel generation: the diffusion approach from image models, applied to text. The architecture differs from the usual autoregressive transformer.
  • 260K context: 260,000 tokens in, up to 65,536 tokens out.
  • Reasoning levels: none, low, medium or high. With nothing set, it reasoned at medium.
  • Tool calling and JSON schema output, both confirmed through FastMetal at launch.
  • Text in and text out only, with no image input.

The empty-answer trap

With no reasoning setting, Mercury 2.5 reasoned, and reasoning tokens are billed as output tokens. That interacts badly with a small max_tokens.

We set max_tokens to 16 in a quick test. The model spent all 16 on reasoning and returned an empty answer. It looks like a broken model, and it is the documented behavior: the budget ran out before any visible text.

For simple conversions or classification, turn reasoning off or set the lowest level:

{ "reasoning": { "enabled": false } }

The same applies to any reasoning model on FastMetal. Our reasoning guide covers how to set it per request.

How fast was it

We ran Mercury 2.5 against Solar Pro 4 on the same 43 prompts. It answered 7.3x faster on average, wrote about half as much text, and still used more output tokens. The full numbers are in Mercury 2.5 vs Solar Pro 4.

Added the same day

We also added Nex AGI's Nex-N2.5-Mini (nex-n2.5-mini-free), a free agentic model aimed at exploring codebases and editing across several files, with a 262,144-token context. It has since been withdrawn upstream as well.

Try Mercury 2.5 right now

Mercury 2.5 is available on FastMetal through one API key. Start in the browser, or call it from the OpenAI SDK.