Mercury 3 tested: the only competitive diffusion LLM, matches GPT-5.6 Luna on coding
volokuleshov · x · 2026-10-21
leolinsky benchmarked Inception AI's Mercury 3, calling it the only lab seriously scaling diffusion architectures for hard text and coding work — and the only competitive general dLLM today.
- Results: slightly below its press release claims, roughly equivalent to GPT-5.6 Luna and Muse Spark 1.3 on average across coding tasks; GScore 52.5 (rank 33) on the GBENCH intelligence benchmark, next to Muse Spark 1.3 and Claude Opus 4.7.
- Value: competitive and near the Pareto front despite a paltry amount of funding vs other neolabs.
- Caveat: it uses more reasoning tokens on hard problems than expected from a diffusion architecture, so speed isn't a major differentiator over GPT-6 Luna or Haiku 5.5; serving specs may be lower than advertised.
- Verdict: 'the hype is real,' author is bullish.
More from Models
- claudefriend screenshot reveals its policy model is a Qwen trained to please Claude — Sauers_ · 2026-10-21
- Microsoft quietly ships Sentiment-Analysis-7B, an RL-tuned Mistral-7B on Hugging Face — NielsRogge · 2026-10-21
- Ask your AI agent its favorite color — it apparently has one — every · 2026-10-21
- Google demos Gemini 4 Argon building a 3D periodic table synced to Tom Lehrer's 'The Elements' — sundarpichai · 2026-10-21
- MiniMax follow-up: M3.1 Flash teased after GTC release wave — MaziyarPanahi · 2026-10-21
- Claude Max 5x limits feel generous: heavy user can't burn through weekly quota — Angaisb_ · 2026-10-21