Mercury 3 challenges GPT-6 Luna: 40 times larger than the first generation, official speed test still exceeds 1,000 tokens per second.

Beating AI News Flash: US AI company Inception releases diffusion language model Mercury 3. Ordinary large models typically generate tokens one by one, while Mercury generates a batch of text in parallel and then repeatedly refines it.

Inception claims that the new model's parameter count is more than 40 times larger than the first-generation Mercury, yet it can still achieve 1,068 tokens per second on NVIDIA GPUs, close to the 1,107 tokens of the previous generation Mercury 2.5. The context has also been expanded from 260,000 to 1 million tokens.

Published self-testing shows that at the medium reasoning tier, Mercury 3's overall score is slightly higher than GPT-6 Luna and Claude Haiku 5.5, with generation speed reaching 4 to 8 times that of the two.

On the DeepSWE programming test, Mercury 3 scored 46.3%, short of Haiku 5.5's 51.6%; switching to SWE-Bench Atlas, it led its rival 51.1% to 32.3%.

Inception's co-founder acknowledged on Hacker News that the new model cannot yet replace Opus or Fable as a primary Agent, but it can already handle subtasks such as code writing.

Mercury 3 is now available through the Inception API, OpenRouter, Baseten, and Vercel. During the launch period, pricing is 50% off: $0.10 per million input tokens and $0.35 for output; regular prices are $0.20 and $0.70 respectively. The official offering includes 100 million free tokens for trial use.

However, ordinary users may not necessarily experience the officially advertised thousand-token speed. OpenRouter shows that Mercury 3's median output speed is about 180 tokens per second, only one-sixth of the official benchmark.