China’s DeepSeek releases V4 Pro 0813, a 1M‑token Mixture‑of‑Experts language model

Chinese AI startup DeepSeek has rolled out DeepSeek V4 Pro 0813, a new large language model built on a Mixture‑of‑Experts (MoE) architecture. The model is designed to handle extremely long inputs, with a stated context window of up to 1 million tokens—enough to ingest very large codebases or document collections in a single pass.

V4 Pro 0813 is being distributed through OpenRouter, an API hub that aggregates access to models from multiple providers. The release follows DeepSeek’s late‑July launch of V4/0731, an open‑source model under the permissive MIT license that, according to community benchmarks, matches leading proprietary systems on coding tasks at roughly 60% lower cost.

The Pro variant is clearly aimed at powering long‑running AI agents and heavy enterprise workloads that require keeping vast amounts of context “in memory.” DeepSeek’s rapid progress—and its willingness to ship competitive models under open licenses—has raised the profile of Chinese labs in Western developer circles. It also adds new pressure to the global debate over how frontier‑level capabilities, pricing, and openness should be balanced against safety and regulatory concerns.