模型發布與開發者API
DeepSeek-V4-Pro Reaches GA: Agentic Benchmarks Improve, API Gains Native Responses Format Support
DeepSeek has moved V4-Pro-0813 into general availability, with the existing `deepseek-v4-pro` name now pointing directly to the new version. The update strengthens coding agents and tool use while adding an OpenAI Responses API compatibility layer, although the published scores still come primarily from the company’s official harness.
On August 13, DeepSeek moved DeepSeek-V4-Pro from preview to general availability across the web, app, and API. Existing applications do not need to change their model identifier: `deepseek-v4-pro` will route directly to the 0813 version. While this in-place upgrade lowers migration costs, it also means teams that require reproducible behavior should immediately freeze their test sets and retain responses and tool traces to avoid mistaking version-related changes for prompt or agent regressions.
The GA update focuses on agentic post-training. DeepSeek reports a Terminal-Bench 2.1 score of 87.9, up 15.8 points from the preview version’s 72.1. NL2Repo rose from 38.5 to 61.5, CyberGym from 52.7 to 83.3, and DeepSWE from 12.8 to 62.7, while Toolathlon-Verified reached 74.1. These benchmarks cover terminal operations, building repositories from specifications, vulnerability reproduction, and tool use. Although the gains are directionally consistent, the agent frameworks, time limits, and reasoning budgets differ across benchmarks, so the results do not yet demonstrate that every production workflow will improve proportionally.
The underlying model remains a Mixture-of-Experts (MoE) architecture with 1.6 trillion total parameters and 49 billion active parameters per token, supporting a one-million-token context window. The technical report describes a hybrid sparse-attention design combining CSA and HCA. At the one-million-token setting, DeepSeek estimates that per-token inference FLOPs are 27% of V3.2’s, while the KV cache is 10% of its size. Post-training first uses SFT and GRPO to develop domain-specific experts, then combines their capabilities through on-policy distillation. The open weights are released under the MIT License. Most MoE experts use FP4, while the remaining components primarily use FP8; full deployment still requires substantial multi-GPU or cluster resources.
The API now also natively supports the OpenAI Responses format and is optimized for integration with Codex-like agents, reducing the need to manually transform streaming tool calls and reasoning fields. Engineering teams should next verify whether tool semantics are fully consistent between Responses and Chat Completions, whether reasoning state persists across long tool-use loops, and whether third parties can reproduce the official GA scores under identical token and time budgets.