Mercury 3 Diffusion LLM Released, Outperforms GPT-6 Luna and Claude Haiku 5.5

Introduction of Mercury 3

Mercury 3, a new diffusion-based large language model, has been introduced. It is presented as the most capable diffusion LLM released to date, showing significant advancements in performance and speed compared to previous models.

Performance and Speed Benchmarks

Across various benchmarks, Mercury 3 reportedly outperforms GPT-6 Luna and Claude Haiku 5.5. The model also operates 4-8 times faster, achieving over 1,000 tokens per second. It maintains this speed despite being more than 40 times larger than its predecessor, Mercury 1.

Technical Specifications and Pricing

Mercury 3 offers a 1M token context window, an increase from Mercury 2.5's 260K tokens. It achieves 1,068 tokens per second on NVIDIA GPUs. The pricing is set at $0.20 per million input tokens and $0.70 per million output tokens, with a launch discount offering 50% off these rates.

Scaling of Diffusion LLMs

The release of Mercury 3 addresses the question of whether diffusion LLMs can scale to compete with autoregressive models for complex use cases. The model has scaled over 40 times in total parameters, with intelligence increasing threefold from Mercury 1 to Mercury 3, suggesting that scaling laws apply to diffusion models similarly to autoregressive models.

Production Use Cases

Early-access partners are already deploying Mercury 3 in production for latency-sensitive applications. These include browser use agents, interactive education platforms, real-time search agents, and customer support systems. For instance, in web-agent tasks, Mercury 3 completed 50% of 50 live Mind2Web navigation tasks, outperforming GPT-6.1-Sol with comparable token volume and agent steps at a lower cost.