Inception says Mercury 3 scales diffusion language models to faster, larger workloads

Inception Labs says Mercury 3, its latest diffusion-based language model, can generate text at more than 1,000 tokens per second while matching or beating some speed-optimized models on benchmarks. The company’s announcement, dated Oct. 19, 2026, describes Mercury 3 as its most capable model to date and says it is available through Inception’s API platform and several partner platforms.

Inception reports a speed of 1,068 tokens per second on NVIDIA GPUs, a context window of one million tokens, and list prices of 20 cents per million input tokens and 70 cents per million output tokens. It says launch pricing cuts those rates in half, with cached input tokens discounted by 90 percent. These are vendor-reported specifications and prices; the announcement does not describe an independent evaluation of the claims.

The model’s significance, if the company’s account holds up in broader testing, is that it applies scaling to a different approach to language generation. Autoregressive models produce text token by token; diffusion models generate and refine text through a process that Inception says can support faster output. The company says Mercury 3 has more than 40 times the parameters of Mercury 1 and that its performance across a range of standard benchmarks has risen threefold. It does not provide the benchmark results or methodology in the announcement.

Inception also reports that Mercury 3 completed half of 50 live Mind2Web navigation tasks in a company-run test, using the same token volume and number of agent steps as a frontier model. It says that result was significantly better than GPT-6.1-Sol, while costing less per token. The test was run in Inception’s own web-agent harness, and the post does not give further details about the comparison or task results.

The launch post includes favorable early feedback from partner companies using Mercury 3 for tutoring, search and customer support. Those accounts describe faster responses and, in some cases, quality comparable to or better than the partners’ previous models. They are customer testimonials rather than a published, independent assessment; Inception does not report a controlled cross-provider evaluation for those uses.

Inception says Mercury 3 is compatible with the OpenAI API and can be accessed through its own platform, OpenRouter, Baseten and Vercel. It also offers enterprise deployments with dedicated capacity, autoscaling and configurable data retention. The company frames the release as evidence that diffusion models can scale to complex commercial applications, a broader claim that will depend on how the model performs in independent tests and real-world deployments.