Mercury 2.5 dLLM hits 1,100 tok/s, 80% off on OpenRouter
Inception's diffusion model claims Haiku-tier quality at $0.04/M input through Sep 7, with tunable reasoning and 260K context for voice and coding agents.
Copy markdown
1,100 tokens/sec, because it denoises in parallel
Mercury 2.5 is a diffusion LLM: it generates blocks of text at once instead of one token at a time. Inception clocks it at ~1,100 tok/s on standard GPUs with 1.38s P50 latency, the kind of speed that makes a streaming response land almost instantly.
$0.04/M input, but the discount clock runs out Sep 7
Through September 7 it's 80% off: $0.04/M input and $0.15/M output (normally $0.20/$0.75), with cache reads at $0.004/M. Even at full price it undercuts most hosted small models, so this week is the cheap window to benchmark it against your current cost floor.
Inception says it jumped a full quality tier
Inception claims a 10+ point intelligence gain over Mercury 2, putting it level with cost-optimized frontier models like Claude Haiku 4.5 and GPT-5.6 Luna. That's a vendor number with no independent benchmarks yet, so treat it as a hypothesis and verify on your own evals before swapping.
Built for the jobs where latency is the product
Tunable reasoning levels, parallel tool calls, and schema-aligned JSON aim it squarely at voice pipelines, search agents, and coding subagents, anywhere a slow first token breaks the UX. It carries a 260K context window and up to 65K output tokens.
How to try it this weekend
It's OpenRouter-exclusive for now and OpenAI-compatible, so it's a one-line base-URL swap into existing code. Point an autocomplete loop or a voice agent at it and the speed difference is immediately obvious; keep it out of sole customer-facing paths while it's still a preview.