Mercury 2.5 pushes diffusion language models toward production use

Inception released Mercury 2.5, which it calls the largest diffusion language model trained so far. Unlike the usual autoregressive models that write one token at a time, a diffusion LLM refines a whole block of text in parallel, and that architecture is what lets Inception quote 1,107 tokens per second on standard NVIDIA GPUs. The company says the new version is about 40 percent more capable than Mercury 2, landing near cost-optimized frontier models like GPT-5.6 Luna (Low) and Claude Haiku 4.5. It has a 260K token context window and, for voice, reaches first token in under 170 milliseconds.

The pitch is speed and price rather than top-of-the-leaderboard quality. Standard rates are $0.20 per million input tokens and $0.75 per million output, with an 80 percent launch discount that drops those to $0.04 and $0.15. Inception points to one customer, OpenCall, cutting P99 latency from several minutes to about one second, though that reflects their particular workload. The announcement leans on benchmarks and metrics and does not open up the model's internals, so the honest read is a fast, cheap model in the Haiku tier rather than a challenger to the strongest reasoning models.

Why it matters

If you build voice assistants, autocomplete, or anything where the user is waiting on each response, sub-170ms first tokens and four-figure throughput change what feels possible. Benchmark it against a small autoregressive model on your own task, since the quality tier is mid-range and the speed claims come from the vendor.