Inception Labs has introduced Mercury 3, which it describes as the most capable diffusion‑based large language model released to date. According to the company’s blog post, Mercury 3 outperforms GPT‑6 Luna and Claude Haiku 5.5 on a range of benchmarks while operating 4–8 times faster, achieving speeds of more than 1,000 tokens per second.
The post states that Mercury 3 demonstrates that text diffusion can scale effectively. It is reported to be more than 40 times larger in total parameters than Mercury 1 and to be vastly more intelligent, yet it retains the same speed as its predecessor. The company notes that intelligence across standard benchmarks has risen threefold from Mercury 1 (mini) to Mercury 3, suggesting that the scaling laws that drove progress in autoregressive models also apply to diffusion models.
Mercury 3 offers a context window of 1 million tokens, an increase from the 260 K tokens available in Mercury 2.5. On NVIDIA GPUs, the model delivers a speed of 1,068 tokens per second. Pricing is set at $0.20 per million input tokens and $0.70 per million output tokens, with a launch discount of 50 % bringing the rates to $0.10 per million input and $0.35 per million output. Coded tokens receive a 90 % discount off the input price.
Early‑access partners are already deploying Mercury 3 in latency‑sensitive applications. In browser‑use agents, the company says Mercury 3 completed 50 % of 50 live Mind2Web navigation tasks, substantially more than GPT‑6.1‑Sol, while using the same token volume and number of agent steps as a frontier model and at a fraction of the per‑token cost. Shuhao Zhang of TinyFish AI observed that the model’s speed allows agents to try more approaches before answering.
For interactive education, Mercury 3 provides near‑real‑time feedback, with hints, marked answers and definitions returning in about a second. Michael Tong of General Learning (RevisionDojo) noted that the quality remains as good as or better than previous models, helping students stay engaged rather than waiting.
In real‑time search agents tested on smart glasses, the company reports that Octen search took roughly 80 ms, with model generation as the bottleneck. Mercury 3 delivered a median generation time of 0.9 seconds and a median end‑to‑end response time of 2.2 seconds, according to Kuan Zou of Octen AI.
For customer support, Mercury 3 is said to beat other models on latency while performing higher on evaluations and in production. Michael Gribben of Lorikeet highlighted that the model keeps responses conversational without sacrificing accuracy for regulated businesses.
The release also introduces companion products: Mercury Voice, tuned for ultra‑low latency in voice applications, and Mercury Decide, a decision model for high‑volume classification that outperforms Jev on benchmarks such as JevBench and LegalBench.
Mercury 3 is available today through the Inception API Platform and via OpenRouter, Baseten, and Vercel. It is OpenAI‑API compatible, allowing drop‑in replacement with no code changes. The company offers 100 million free tokens for trial and provides enterprise evaluations that include workload fit, assessment design, and performance under serving constraints.