Inception Launches Mercury 3: Diffusion LLM Exceeds 1,000 Tokens/Second With 1M Context

What happened

According to Inception's official press release published on BusinessWire on October 19, Mercury 3 is the company's fourth model release since Mercury 1 launched in June 2025: Mercury 2 added reasoning to diffusion models, Mercury 2.5 introduced tunable reasoning with a 260,000-token context window, and Mercury 3 is described by the company as the largest intelligence jump of any Mercury release.

The company says Mercury 3 is more than 40 times the size of Mercury 1 yet runs at comparable speeds — over 1,000 tokens per second on widely available NVIDIA GPUs. Traditional autoregressive models generate text one token at a time, while diffusion models can generate multiple tokens in parallel; Inception says Mercury 3 shows diffusion models keep that parallel-generation speed advantage as they scale, with the limits coming from compute and data rather than the algorithm itself.

Mercury 3 is available starting today through the Inception API, Baseten, OpenRouter and Vercel; it is OpenAI API compatible, so existing users can switch without rewriting code. Enterprise deployments support dedicated capacity, autoscaling, compliance controls and configurable data retention. Early-access customers are already testing it in browser-use agents, interactive education, real-time search agents and customer support.

Key facts

1. Speed:over 1,000 tokens per second on widely available NVIDIA GPUs; the company says quality matches or exceeds GPT-6 Luna and Claude Haiku 5.5 (all company-reported benchmark figures).
2. Context window:1 million tokens, up from 260,000 in Mercury 2.5.
3. Pricing:$0.20 per million input tokens and $0.70 per million output tokens; 50% off at launch ($0.10 input, $0.35 output), with cached tokens at 90% off the input price.
4. Early customers: are already using it in browser-use agents, interactive education (General Learning/RevisionDojo says student feedback returns in about a second), real-time search agents and customer support.
5. Company background:Inception is based in Redwood City, California, and describes itself as the first company to commercialize diffusion large language models; backers include Menlo Ventures, Mayfield, Innovation Endeavors, M12 (Microsoft's venture fund), Snowflake Ventures, Databricks Ventures, and individual investors such as Andrew Ng, Andrej Karpathy and Eric Schmidt.