- Inception released Mercury 3, a diffusion-based language model the company says runs at 1,068 tokens per second with a 1 million-token context window.
- On Inception’s own benchmark table, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests, including SWE-Bench Atlas and IFBench.
- Haiku 5.5 still wins on DeepSWE v1 and MRCR v2, Inception’s benchmarks show.
- Launch pricing is $0.10 per 1M input tokens and $0.35 per 1M output tokens, rising to $0.20 / $0.70 after the promotional period.
What Happened
AI startup Inception has released Mercury 3, the latest version of its diffusion language model, which generates text by refining a full draft answer in parallel rather than one token at a time, Marktechpost reported (Oct 20, 2026). Inception says the model reaches 1,068 tokens per second on Nvidia GPUs with a 1 million-token context window, and has more than 40 times the parameters of the original Mercury model.
Why It Matters
Diffusion LLMs have been pitched as a faster alternative to autoregressive models since fewer sequential steps are needed to produce an answer. Mercury 3 is Inception’s attempt to show that approach holds up at larger scale: the company says average benchmark quality rose roughly 3x between Mercury 1 and Mercury 3 while speed stayed above 1,000 tokens per second.
Technical Details
On Inception’s self-reported benchmark table, run at medium settings, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests. Its largest margins are on SWE-Bench Atlas (51.08, versus 33.33 for Luna and 32.26 for Haiku 5.5) and IFBench (76.84, 15 points above Luna). Haiku 5.5 still wins on DeepSWE v1 (51.62 vs. 46.31) and MRCR v2 (78.40 vs. 75.50), with Luna slightly ahead on AA-Omniscience-Public and AA-LCR v1.1 — the benchmarks suggest long-context retrieval and longer coding tasks remain Mercury 3’s weaker areas.
Who’s Affected
The model is available via the Inception API, OpenRouter, Baseten, and Vercel, with an OpenAI-compatible API and no open weights. Inception quoted four early-access partners, including TinyFish and Lorikeet, on latency and task-completion results, but these are customer-supplied figures rather than independent evaluations.
What’s Next
Inception also launched two sibling models, Mercury Voice for voice-agent latency and Mercury Decide for classification tasks. For buyers comparing speed-focused models against incumbents, MegaOne AI’s own head-to-head testing of diffusion-based generative models offers a related framework for weighing speed claims against output quality.
