DeepSeek releases V4-Pro-0813 with DSpark speculative decoding and stronger agentic benchmarks

DeepSeek released DeepSeek-V4-Pro-0813, the official version superseding the V4-Pro preview, built on the same model structure with an added DSpark speculative decoding module. The company reports gains over the preview on benchmarks including Terminal Bench 2.1 (87.9 vs 72.1), DeepSWE (62.7 vs 12.8) and NL2Repo (61.5 vs 38.5), and says the model is broadly competitive with the strongest proprietary models. The release ships without a Jinja chat template, instead providing an encoding folder, and adds low, high and max levels to the reasoning_effort parameter; weights are MIT licensed and vLLM and SGLang launch commands are documented.