Inception launches Mercury 3, a diffusion-based LLM that it claims outperforms GPT-6 Luna, Claude Haiku 5.5, and GLM-5.3 Flash while being 4x to 8x faster

AI model inference gets up to 8x faster and cheaper, cutting response times for applications.

Inception has launched Mercury 3, a new diffusion-based large language model (LLM) that it claims outperforms leading models like GPT-6 Luna, Claude Haiku 5.5, and GLM-5.3 Flash. Unlike traditional LLMs that generate text token by token sequentially, Mercury 3 utilizes parallelized inference to generate and refine multiple tokens simultaneously, a technique common in image and video generation but novel for text. This approach allows Mercury 3 to operate significantly faster, with Inception reporting speeds up to 1,000 tokens per second and a 4x to 8x performance improvement over competitors on agentic benchmarks. The company also highlighted cost efficiencies, with launch pricing at $0.10 per million input tokens and $0.35 per million output tokens, representing a discount on output compared to rivals. Inception also employed recursive self-improvement using swarms of Mercury 3 agents, boosting the model's speed by an additional 16%.

Mercury 3's core innovation lies in applying diffusion models, typically used for image generation, to LLMs. This architectural shift enables parallelized inference, drastically reducing processing time compared to the autoregressive nature of traditional LLMs. The claimed speed gains and performance parity with established models challenge the industry's prevailing strategy of simply scaling up parameter counts, suggesting a new path toward more efficient and responsive AI. This could signal a broader shift in LLM development, prioritizing architectural innovation for speed and cost-effectiveness over brute-force scaling.

The implications for AI economics are substantial, as faster and cheaper inference directly translates to lower operational costs for businesses deploying AI. This approach may also democratize access to high-performance AI by reducing the computational resources required. The development could also spur further investment and research into diffusion-based LLMs, potentially accelerating the timeline for their widespread adoption and influencing the competitive landscape against established players like OpenAI and Anthropic.