Inception's argument has always been that diffusion language models can be made to scale, and Mercury 3 is the post where the company says so outright — "Mercury 3 answers that question: they can", closing on "Diffusion scales." It is pitched as the most capable diffusion LLM released to date: more than 40× larger than Mercury 1 in total parameters, with a 1M-token context (260K in Mercury 2.5) and 1,068 tokens per second on NVIDIA GPUs. Exact parameter counts are not disclosed, as with the rest of the line.

The speed-and-price case is the concrete one. $0.20 / $0.70 per 1M tokens, currently half that on a launch promotion ($0.10 / $0.35, which is what OpenRouter is serving), with cached input 90% off. Partner numbers do more work than the benchmarks: Octen AI reports a 0.9s median generation time and 2.2s end-to-end for real-time search, in which the model contributes roughly 80ms. On 50 live Mind2Web navigation tasks it completed 50%, which Inception says beats GPT-6.1-Sol. Available through Inception's API, OpenRouter, Baseten and Vercel, OpenAI-compatible, with 100M free tokens at launch.

The intelligence case is asserted rather than shown, and the one independent measure runs the other way. The post claims intelligence "has risen 3x from Mercury 1" across "a range of standard benchmarks" without naming them, and that Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 — again with no table. The three benchmarks named anywhere in the post are JevBench, LegalBench and Mind2Web, and only Mind2Web carries a figure. Meanwhile Artificial Analysis has the line at 14 for Mercury 2 and 12 for Mercury 2.5 on Intelligence Index v4.3.2 — the previous generation went down — against Haiku 5.5's 43. AA has not scored Mercury 3 at filing, so treat the comparison claims as the vendor's until it does.

Two siblings ship alongside. Mercury Voice is tuned for the latency floor voice agents need, and Mercury Decide is a high-volume classification model that Inception says "outperforms Jev on benchmarks including JevBench and LegalBench" — no scores given. Decide puts Inception into the same emerging decision-model category as Liquid's d1, from the opposite direction: d1 is open-weight and runs on a laptop, Decide is closed and runs at a thousand tokens a second. No weights for any of them.

Model Details

Context window 1,000,000
License Proprietary

Benchmark Scores

Benchmark Score Mode
Mind2Web (50 live navigation tasks) 50% —

Variants

Name Parameters Notes
Mercury 3 — flagship dLLM; 1M context, 1,068 tok/s on NVIDIA GPUs, $0.20/$0.70 per 1M (launch promo $0.10/$0.35)
Mercury Voice — tuned for the low-latency requirements of voice agents
Mercury Decide — high-volume classification; claimed to outperform Jev on JevBench and LegalBench
diffusionefficiencyreasoningagentsproprietary

Related