Mercury 2.5 diffusion LLM hits 1,107 tok/s, launch price 80% off
Inception's parallel-decoding model matches Haiku-class quality with 260K context and sub-170ms voice—while open Nex-N2.5 models land free on OpenRouter.
Copy markdown
1,107 tokens/sec, because it decodes in parallel
Mercury 2.5 is a diffusion LLM that refines many tokens at once instead of one at a time, hitting 1,107 tok/s on standard NVIDIA GPUs—Inception says no autoregressive model at this price comes close. For you: streaming UIs and agent loops that feel instant instead of typewritered.
The price: $0.20/$0.75, and 80% off right now
Standard rates are $0.20 per million input and $0.75 output; a launch discount cuts that to $0.04/$0.15. The Inception API also hands you 100M free tokens to start, and it's already live on OpenRouter and Baseten.
40% smarter than Mercury 2, with a 260K window
Inception claims a 40% intelligence jump, putting Mercury 2.5 in the same tier as GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5—now with a 260K context and up to 65K output. Tunable reasoning, parallel tool calls, and schema-aligned JSON come built in.
Build this weekend: a voice agent that answers in 170ms
Mercury Voice posts time-to-first-token under 170ms, and one deployment (OpenCall) ran production phone calls at ~170ms median latency. That's past the threshold where a caller notices the pause—wire it to a telephony API and ship a real-time agent.
For coders: context compaction 82% faster, 90% cheaper
Augment Code moved history compaction to Mercury and cut latency 82% (roughly 150s down to 27s) and cost 90%, with no quality regression. If your agent stalls every time it compresses a long session, this is close to a drop-in fix.
Elsewhere: Nex-N2.5 lands free and open on OpenRouter
Nex-AGI open-sourced its Nex-N2.5 line (Mini, Pro, Max) and put the Mini and Pro tiers up free on OpenRouter—another cheap, self-hostable backend for agents worth benchmarking against your current default.