Inception Launches Mercury 3: A High-Speed Diffusion Language Model
Mercury 3 leverages diffusion architecture to achieve rapid text generation and advanced reasoning capabilities.
1Mercury 3 Model Overview
On October 19, 2026, Inception released Mercury 3, a powerful diffusion-based language model. It features a massive 1 million token context window and supports a maximum response length of 131,072 tokens. The model is designed for high-speed performance, reportedly generating over 1,000 tokens per second on NVIDIA GPU hardware.
2Diffusion vs Autoregressive Models
Unlike traditional autoregressive models that generate tokens sequentially, Mercury 3 utilizes a diffusion approach. It creates a draft of the entire response and refines it through multiple parallel passes. This mechanism significantly reduces the time required for long-form generation, as the model does not wait for each preceding token to complete.
3Performance and Benchmark Results
Inception reports that Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 in 8 out of 12 standard benchmarks. Key metrics include IFBench at 76.84, GPQA Diamond at 89.23, and SWE-Bench Atlas at 51.08. However, the model trails behind Claude Haiku 5.5 in specific tests like DeepSWE v1 and MRCR v2.
4Independent Real-World Metrics
Independent data from OpenRouter indicates a median generation speed of approximately 243 tokens per second, lower than Inception's lab-based claims. On the Terminal-Bench 4.0 test, the model scored 3%, highlighting potential challenges in complex, multi-step terminal tasks compared to its strong performance in shorter, instruction-following scenarios and agentic programming workflows.
5Inception Startup Background
Based in Redwood City, California, Inception is led by CEO Stefano Ermon, a Stanford professor, and Chief Scientist Vladimir Kuleshov. The company focuses on commercializing diffusion language models. Mercury 3 represents their fourth major release, following a series of experimental models that have progressively improved in reasoning depth and context handling capabilities.
6Technical Scaling of Diffusion
Inception claims that diffusion scales effectively with model size. Mercury 3 is over 40 times larger than the original Mercury 1 model while maintaining high generation speeds. The development team suggests that the benefits of parallel generation become more pronounced as the length of the required reasoning and output increases.
7Limitations and Constraints
Mercury 3 is strictly a text-based model and lacks the multimodal capabilities found in GPT-6 Luna or Claude Haiku 5.5. It cannot process images, screenshots, or documents. Furthermore, the non-linear nature of its diffusion-based generation makes the internal reasoning process more difficult to interpret compared to standard autoregressive LLM architectures.
8Inception Chat Interface Features
The Inception Chat platform provides a visual demonstration of diffusion through the 'Diffusion Effect' toggle. Users can observe the model refining its output from noise to coherent text. The interface includes specialized playgrounds for voice interaction, web-based search, decision-making tasks, and coding, allowing for immediate testing of the model's agentic capabilities.
9Practical Coding with Mercury 3
The Code Playground allows users to build browser-based games like Asteroids in seconds. By providing a prompt, the agent reads existing files, writes new logic, and verifies the code. In tests, the model successfully implemented features like particle explosions, shield power-ups, and enemy UFO behavior within a very short timeframe.
10Live Design Editor Functionality
Beyond coding, Mercury 3 serves as a design assistant in the Live Design Editor. Users can manipulate web layouts, such as a coffee shop landing page, using natural language commands. The model updates the canvas in real-time, allowing for rapid iteration of UI elements like button colors, themes, and card layouts.
11Availability and Access
Mercury 3 is currently available via Inception Chat, the Inception API, and third-party aggregators including OpenRouter, Baseten, and Vercel. Users should note that older versions, Mercury 2 and 2.5, are no longer accepting new users. Existing clients of these legacy models will retain access only until the cutoff date of November 15, 2026.
12Strategic Implications for AI
The release of Mercury 3 marks a significant step for diffusion-based language models in commercial applications. By proving that diffusion can scale for complex reasoning, Inception challenges the dominance of autoregressive architectures. Future developments will likely focus on improving performance in long-horizon tasks and expanding the model's utility in specialized agentic environments.
Original article: INCRYPTED — Inception Launches Mercury 3: A High-Speed Diffusion Language Model