Inception launches Mercury 3, a diffusion-based LLM it claims is "comparable to Luna and Haiku models" but 4x to 8x faster, generating ~1,000 tokens per second

LLM inference speed across all models improves by up to 8x, reducing latency and cost for real-time AI applications.

Inception has launched Mercury 3, a new diffusion-based large language model (LLM) that claims performance comparable to Anthropic's Luna and OpenAI's Haiku models but with significantly increased speed. The company states Mercury 3 can generate approximately 1,000 tokens per second, which is 4x to 8x faster than its competitors. This diffusion LLM architecture, previously used primarily for image and video generation, parallelizes inference to generate and refine multiple tokens simultaneously, unlike traditional autoregressive LLMs that process tokens sequentially. Inception also reported that Mercury 3 agents achieved a 16% speed increase through recursive self-improvement. The model is currently priced at $0.10 per million input tokens and $0.35 per million output tokens, with the output token price representing a 35% discount compared to leading models. This launch price is 50% off the eventual standard price, and Inception is exploring options to make the promotional pricing permanent.

Inception's Mercury 3 represents a significant challenge to the dominant paradigm in LLM development, which heavily relies on scaling up traditional transformer architectures and investing in massive compute infrastructure. By successfully applying diffusion model technology to text generation, Inception demonstrates a viable alternative path that prioritizes speed and efficiency. This could force other major AI labs, including Google DeepMind which is also exploring diffusion LLMs, to re-evaluate their R&D strategies and potentially shift focus from sheer model size to architectural innovations that deliver comparable or superior performance at lower costs.

The breakthrough in diffusion LLMs, if broadly applicable, has the potential to democratize advanced AI capabilities by making them more accessible and affordable. Faster response times and lower operational costs could enable a wider range of real-time AI applications, from interactive chatbots and virtual assistants to more complex agentic systems that require rapid iteration. The key question remains whether Mercury 3's speed advantage will translate into long-term competitiveness against established frontier models or remain a specialized solution for specific use cases.