Inception Launches Mercury 2, First Diffusion-Based Language Reasoning Model

Inception Labs, a pioneering AI startup, has unveiled the industry's first diffusion-based reasoning artificial intelligence model, combining high speed with cost efficiency.

Unlike traditional language models that generate text token by token, Mercury 2 refines multiple text blocks simultaneously. The startup compares this process to an editor revising an entire draft at once, rather than correcting individual words sequentially.

Speed and pricing stand out as the primary advantages. According to Inception Labs, Mercury 2 achieves 1,009 tokens per second on Nvidia Blackwell GPUs, with an end-to-end latency of just 1.7 seconds. In contrast, Gemini 3 Flash records 14.4 seconds, and Claude Haiku 4.5 reaches 23.4 seconds when reasoning features are enabled. The company asserts that its output quality rivals leading speed-optimized models.

Pricing is set at $0.25 per million input tokens and $0.75 per million output tokens. This represents a 50% reduction in input costs and a fourfold decrease in output costs compared to Gemini 3 Flash ($0.50/$3.00). Relative to Claude Haiku 4.5 ($1.00/$5.00), input pricing is approximately one-fourth, and output pricing is less than half.

The model supports a 128K context window, tool usage, and JSON output. Inception Labs targets companies managing latency-sensitive applications, such as voice assistants, coding tools, and search systems.

Mercury 2 is now available via an OpenAI-compatible API. Companies can apply for early access, and the model can be tested directly within the chat interface.

The AI Industry Seeks the Future Beyond Transformers

Last November, Inception secured $50 million in funding from investors including Microsoft, Nvidia, and Snowflake. The startup showcased its initial prototype in early 2025. With the launch of Mercury 2, Inception has now released a production-ready model equipped with reasoning capabilities.

Google DeepMind is also developing diffusion-based language models. Gemini Diffusion performed comparably to the Gemini 2.0 Flash Lite model in benchmarks. However, since first demonstrating the diffusion experiment in May 2025, Google has made no further announcements regarding the project.

Broadly speaking, an increasing number of startups are exploring alternatives to the mainstream Transformer architecture. Whether diffusion-based language models can sustain long-term viability remains to be seen.