Inception releases Mercury 3, a diffusion language model it says beats GPT 6 Luna and Claude Haiku 5.5 on quality at over 1,000 tokens a second
1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.
Context
If the company's numbers hold up, Mercury 3 shows that diffusion language models can scale to compete with standard models on quality while staying several times faster and cheaper.
The benchmark, speed and quality claims are self reported by Inception and have not been independently replicated. Stefano Ermon is a Stanford professor and Inception cofounder, and Inception builds diffusion language models, which refine a whole response in parallel instead of writing one token at a time.
Deep dive
On October 19, Stefano Ermon posted that Inception is releasing Mercury 3, which he called by far its most powerful model and the most intelligent diffusion LLM ever released, and said it outperforms GPT 6 Luna and Haiku 5.5 on quality while running at over 1,000 tokens per second.
Inception's launch post says Mercury 3 is on par with or better than speed optimized models such as GPT 6 Luna and Claude Haiku 5.5 across a range of benchmarks while running 4 to 8 times faster, at 1,068 tokens per second on Nvidia GPUs. It says the model is more than 40 times larger than Mercury 1, that the scaling laws behind standard models hold for diffusion too, and that its context window is now 1 million tokens, up from 260,000 for Mercury 2.5.
The listed price is $0.20 per million input tokens and $0.70 per million output tokens, cut by half at launch to $0.10 and $0.35. Mercury 3 is available through Inception's API, OpenRouter, Baseten and Vercel, and is compatible with the OpenAI API. Inception's model page says Mercury 2 and Mercury 2.5 are no longer offered to new users.
Startup Fortune notes that every one of the performance numbers comes from Inception itself and has not been replicated by an independent evaluator.
What to expect next
Watch independent evaluations from groups such as Artificial Analysis and OpenRouter usage data, which will show whether the quality claims against GPT 6 Luna and Claude Haiku 5.5 hold on real workloads. The bigger question is whether diffusion keeps scaling: if Inception can push the approach toward frontier quality without losing its speed, larger labs may adopt it. Also watch what happens to pricing after the launch discount ends and whether Mercury 2 users migrate before access ends in November.