A Diffusion Model That Writes 1,100 Tokens a Second

Most language models write the way people read, one token after another from left to right. Inception has spent two years arguing that this is not the only option, and its new release makes the case louder. Mercury 2.5, announced this week, generates text at about 1,107 tokens a second on ordinary NVIDIA hardware, a rate that leaves conventional models well behind.

The speed comes from the underlying method. Rather than predicting the next word in sequence, a diffusion model starts with rough, noisy output and refines the whole passage in parallel over several passes, the same broad idea that powers most AI image generators. Because it works on the text as a block instead of a chain, it can produce long stretches far faster. Inception says Mercury 2.5 is the largest diffusion language model trained to date.

Fast is only half the story

Raw speed is easy to dismiss if the writing is poor, so the more interesting claim is about quality. Inception puts Mercury 2.5 roughly level with the cost-efficient tier of the frontier labs, naming GPT-5.6 Luna Low, Gemini 3.5 Flash-Lite and Claude Haiku 4.5 as comparison points. The company also reports a 40 percent jump in capability over the first Mercury, along with a 260,000-token context window, tunable reasoning, parallel tool calls and structured JSON output. Those are the features that decide whether a model can sit inside real software rather than just answer questions.

The pricing is aimed squarely at that market. Standard rates are $0.20 for a million input tokens and $0.75 for a million output tokens, with a launch discount that cuts those to $0.04 and $0.15 for now. For workloads that lean on speed, such as autocomplete, live agents or anything a person waits on, the combination of low latency and low price is the pitch.

A word of caution is fair here. Vendor benchmarks flatter the vendor, and Mercury 2.5 is being measured against the cheaper, faster models rather than the top-end systems from Google, OpenAI or Anthropic. Diffusion for text also remains the road less travelled, and the wider industry has not yet settled on whether its speed advantage holds across every kind of task. What Mercury 2.5 does establish is that the approach is no longer a curiosity. It is fast, it is cheap, and it is good enough to compete.