DeepSeek V4 Pro (0813)

Overview. DeepSeek-V4-Pro-0813 is the general-availability release of DeepSeek's V4 Pro line, succeeding the V4-Pro Preview. It's a sparse mixture-of-experts model with roughly 1.6T total parameters and 49B active per token, released under an MIT license with open weights published on HuggingFace. Developers reach for it when they need a reasoning-focused model that can hold a full million-token context in a single call.

Key strengths. The model is built as a dedicated reasoning system, with selectable reasoning-effort levels (low, high, max) that expose visible thinking traces during generation. Its headline capacity numbers stand out: a 1,048,576-token context window paired with up to 384K output tokens, letting it work through very large inputs and long, multi-step outputs in one pass. On the Artificial Analysis Intelligence Index it scores 42 at max reasoning effort — good for #6 of 112 tracked models and well above the median for open-weight models of similar size (median: 22).

When to use it. Use DeepSeek-V4-Pro-0813 for long-document and long-codebase analysis, multi-step reasoning chains, or agentic workflows where the value of a huge context window and adjustable reasoning depth outweighs the need for instant responses. Its generation profile — extended thinking traces and large output budgets — fits batch and asynchronous pipelines better than latency-sensitive, interactive chat. Dial reasoning effort down to low for faster, lighter passes, or up to max when task complexity justifies the extra thinking time.

Limitations. The model is text-only: it has no vision, audio, or native document-file ingestion, so image- or speech-driven use cases need a different model. Its batch-oriented latency profile and extended thinking traces mean it isn't a fit for realtime, low-latency applications, and no training data cutoff date has been published for it. Available benchmark data is currently limited to a single aggregate Intelligence Index score, without confirmed per-domain (code, math, agentic, etc.) breakdowns.

Sovereignty note. DeepSeek-V4-Pro-0813 is released under an MIT license with open weights, so it can be self-hosted on infrastructure you control rather than relying solely on a third-party API. No data-residency or CLOUD Act exposure details for hosted endpoints of this model have been confirmed yet.