Mercury 2: The First Reasoning Diffusion Language Model (1,000+ tokens/sec)

About this video

Mercury Two: The First Reasoning Diffusion LLM (1,000+ tokens/sec) - Speed Without Sacrificing Quality Inception Labs releases Mercury Two, a reasoning diffusion-based LLM that exceeds 1,000 tokens per second by generating multiple tokens per forward pass and iteratively refining output, rather than using autoregressive token-by-token generation. The script compares its throughput to Haiku (~89 t/s) and GPT-5 Mini (~71 t/s) and argues diffusion provides built-in error correction that can improve reasoning. Mercury Two is presented as maintaining quality while being fast, tying GPT-5 Mini on AIME 2025 at 91.1 and scoring competitively on GPQA and LiveCodeBench. A demo shows Mercury Two versus Haiku with selectable reasoning levels (instant/low/medium/high) and an agentic workflow that uses browser tool calls to find and summarize AI-related Hacker News stories and comments, emphasizing reduced latency in tool-heavy loops. The model supports tool use, structured outputs, RAG, and a 128k context window, and is priced at $0.25 per million input tokens and $0.75 per million output tokens. The script notes an OpenAI-compatible API (swap base URL/model string/API key) and mentions the demo uses Vercel's AI SDK, with code to be linked in the video description. It contrasts industry efforts focused on incremental autoregressive inference optimizations with Mercury Two's model-level approach, highlighting latency-sensitive use cases like voice interfaces, coding iteration, and chat apps, and encourages viewers to try the API platform and playground. 🔗 Try Mercury 2 API Platform: http://platform.inceptionlabs.ai/ Playground: https://chat.inceptionlabs.ai/ Inception is a Palo Alto-based AI lab founded by researchers from Stanford, UCLA, and Cornell - including Stefano Ermon, co-inventor of the diffusion methods behind modern image and video generation. Backed by Menlo Ventures, M12 (Microsoft), NVentures (NVIDIA), Databricks, and individual investors including Andrew Ng, Andrej Karpathy, and Eric Schmidt. 💻 Repo to Demo App Coming soon! 00:00 Mercury Two Breakthrough 00:20 Why Speed Used to Cost Quality 00:43 Diffusion Reasoning Explained 01:50 Speed and Benchmark Results 02:16 Live Demo Versus Haiku 02:40 Agentic Tool Use Example 03:40 API Setup and Pricing 04:36 Best Use Cases for Low Latency 05:16 Diffusion vs Autoregressive This video is sponsored by Inception Labs. 06:11 Industry Race and Big Picture 07:14 Wrap Up and Try the API