Mercury 2.5: a 1,107 tok/s diffusion LLM at Haiku-tier quality
The largest diffusion model yet ships on OpenRouter and Baseten at $0.04/M for launch, with 100M free tokens — plus Mercury Voice and open Nex-N2.5 coders.
Copy markdown
1,107 tokens/second, and it's the biggest diffusion LLM yet
Mercury 2.5 runs at 1,107 tok/s on widely available NVIDIA GPUs — Inception calls it the largest diffusion language model ever trained — while gaining 40% in intelligence over Mercury 2. It now sits in the tier of Claude Haiku 4.5, Gemini 3.5 Flash-Lite, and GPT-5.6 Luna (Low), with a 260K context and tunable reasoning.
The price: $0.04/M at launch, plus 100M free tokens
Standard rates are $0.20 in / $0.75 out per million tokens, cut 80% to $0.04 / $0.15 for launch, with 100M free tokens to test. That's cheap enough to point high-volume RAG, classification, and extraction traffic at it today via Inception's API, Baseten, or OpenRouter.
Build this weekend: sub-second voice, typing-speed coding
In production, OpenCall cut P99 latency from minutes to about one second (~170ms median) and Augment Code dropped context-compaction time 82%, from ~150s to 27s. With parallel tool calls and schema-aligned JSON built in, you can wire up a real-time voice agent or an autocomplete loop that keeps pace with typing.
Mercury Voice and Mercury Router enter preview
Alongside 2.5, Inception previewed Mercury Voice, a speech-native diffusion model aimed at low-latency voice agents, and Mercury Router, which sends each request to the cheapest model that can handle it — the routing layer latency-sensitive apps usually have to build themselves.
Elsewhere: Nex-N2.5 agentic coders go open and free
Nex AGI put two Apache-2.0 agentic-coding models on OpenRouter and Hugging Face — Nex-N2.5-Pro (262K context) and Nex-N2.5-Mini (35.1B params) — both with tool calling and adjustable reasoning, free for a limited time. The weights are downloadable, so you can self-host whichever one fits your GPU.