On October 19, 2026, AI developer Inception Labs announced Mercury 3, describing it as the largest and most capable diffusion-based language model (Diffusion LLM) built to date. The model departs from the standard autoregressive paradigm that dominates modern natural language processing, aiming to prove that diffusion methods can scale effectively for text generation. Compared to its first-generation predecessor, Mercury 1, the new system expands model scale by more than fortyfold.
Throughput and context capacity form the core of the announcement. Running on Nvidia graphics hardware, Mercury 3 achieves generation speeds of up to 1,068 tokens per second while supporting a context window of one million tokens, up from 260,000 tokens in version 2.5. According to company benchmarks, the architecture operates between four and eight times faster in latency and raw throughput than autoregressive counterparts such as Claude Haiku 5.5 and GPT-6 Luna.
Inception Labs is making Mercury 3 available via an application programming interface priced at 0.10 US dollars per million input tokens and 0.35 US dollars per million output tokens. Alongside the flagship model, the company introduced two dedicated variants: Mercury Decide, an engine optimized for fast decision-making workflows with a reported latency of 160 milliseconds, and Mercury Voice, tailored for voice interactions.
The launch aligns with a broader industry push toward high-throughput inference across expansive context windows. On the same day, Chinese AI lab MiniMax introduced MiniMax M3.1 Flash, a mixture-of-experts model featuring 428 billion total parameters with approximately 23 billion active parameters. Leveraging sparse attention techniques, that model also processes up to one million context tokens, underscoring how engineering focus across the sector is shifting from sheer parameter growth toward operational efficiency.
For teams developing autonomous software agents and retrieval pipelines, raw inference speed directly dictates viability. Multi-step agent workflows require models to ingest and evaluate extensive background documentation rapidly. Higher token generation rates substantially reduce cycle times, making complex background reasoning far more practical in production environments.