Inception Releases Mercury 3 Diffusion LLM With 1M Context, Claiming 1,068 Tokens per Second

Mercury 3 is Inception's fourth Mercury release after Mercury 1, Mercury 2 and Mercury 2.5. Inception did not disclose the parameter count but said the model is more than 40 times larger than Mercury 1 and that average quality across standard benchmarks rose about threefold between the two models while speed stayed above 1,000 tokens per second. Mercury 1 launched in 2025; Mercury 2.5 added tunable reasoning and a 260,000-token context window.

Unlike autoregressive models that generate one token after another, a diffusion LLM starts with a rough draft of the whole answer and refines every position in parallel over a few passes. Inception is presenting Mercury 3 as evidence that the approach scales as models grow larger. The company said the model runs on NVIDIA GPUs served by Inception but did not identify the GPU. It disclosed no open weights or quantized builds.

All benchmark scores cited by Inception came from its own table, with each model run at medium settings. Mercury 3 reached 76.84 on IFBench, 15 points above GPT-6 Luna, and 51.08 on SWE-Bench Atlas, against 33.33 for Luna and 32.26 for Haiku 5.5. It scored 89.23 on GPQA Diamond. Haiku 5.5 won DeepSWE v1 at 51.62 versus 46.31 for Mercury 3 and MRCR v2 at 78.40 versus 75.50. GPT-6 Luna edged Mercury 3 on AA-Omniscience-Public, 46.94 to 46.44, and on AA-LCR v1.1, 78.30 to 77.67. Inception identified long-context retrieval and longer coding tasks as weak spots.

Inception quoted four early-access partners, whose statements were provided by the company and were not independent evaluations. TinyFish ran 50 live Mind2Web web-navigation tasks and said Mercury 3 completed 50 percent; Inception said that beat GPT-6.1-Sol but did not give a Sol number. Octen AI reported 0.9 seconds median generation and 2.2 seconds end to end for smart-glasses search. RevisionDojo said hints and marked answers return in about a second. Lorikeet said Mercury 3 beat other models on latency and evaluations for support agents.

Mercury 3 is available on the Inception API Platform, OpenRouter, Baseten and Vercel. The API is OpenAI-compatible and uses the model string mercury-3 and a reasoning_effort parameter. List pricing is $0.20 per 1 million input tokens and $0.70 per 1 million output tokens, with a 50 percent launch discount to $0.10 and $0.35. Cached input is 90 percent off, and new accounts receive 100 million free tokens. Enterprise plans add dedicated capacity, autoscaling and configurable data retention. Inception also released Mercury Voice, aimed at time to first answer token in voice agents, and Mercury Decide, a classification model.