DeepSeek just quietly shipped the general-availability release of DeepSeek V4 Pro 0813, its flagship reasoning model, with no blog post, no press release, and no changelog entry. The API endpoint simply started pointing to the new build. What it did ship, though, is a dramatic improvement in agentic and coding performance, open weights under the MIT license, and a price increase that changes the cost calculus for anyone running it at scale.

The architecture hasn't changed. The post-training has.

DeepSeek V4 Pro 0813 is the official release of DeepSeek V4 Pro, superseding the preview version, with greatly enhanced agentic capabilities. It is built on the DeepSeek V4 Pro Preview model structure, with a DSpark speculative decoding module attached. DeepSeek attributes these gains entirely to improved post-training , the architecture of DeepSeek V4 Pro 0813 is unchanged from the preview.

Under the hood, the model has 1.6 trillion total parameters with roughly 49 billion active per token, pretrained on more than 32 trillion tokens. It uses a hybrid attention system designed to cut inference costs at long context lengths, supporting a 1,048,576-token (1M) context window and up to 384,000 output tokens. The architecture combines two novel attention mechanisms from the DeepSeek-V4 technical paper:

• Compressed Sparse Attention (CSA): compresses the KV cache along the sequence dimension, then applies sparse attention so each query only attends to a subset of compressed entries. This dramatically reduces memory at long contexts.
• Heavily Compressed Attention (HCA): applies even more aggressive compression for a smaller memory footprint, at the cost of some resolution.

This hybrid attention stack reaches 27% of V3.2's per-token inference FLOPs and 10% of V3.2's KV cache at 1M context. The model was pre-trained on 32T+ tokens with the Muon optimizer for faster convergence; post-training is a two-stage pipeline of domain-specific expert cultivation followed by unified consolidation via on-policy distillation. On-policy distillation means the student model learns from the output distributions of multiple specialist teacher models on its own generated trajectories, rather than from static labeled data.