From the source
Summary by forck
Inception released Mercury 3, a diffusion language model that writes answers in parallel rather than token by token, reporting 1,068 tokens per second on
NVIDIA GPUs and a 1M-token context window.
- The company says Mercury 3 is its fourth Mercury release and more than 40x larger than Mercury 1, with average quality across standard benchmarks rising about 3x between the two models while speed stayed above 1,000 tokens per second.
- On Inception's own benchmark table, run at medium settings, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests, including SWE-Bench Atlas at 51.08 against 33.33 and 32.26, and IFBench at 76.84.
- It loses to Haiku 5.5 on DeepSWE v1 (46.31 versus 51.62) and MRCR v2 (75.50 versus 78.40), leaving long-context retrieval and longer coding tasks as weak spots.
- Mercury 3 is available through the Inception API Platform, OpenRouter,
Baseten and Vercel, with an -compatible API using the model string mercury-3 and a reasoning_effort parameter.