DeepSeek released an upgraded version of its V4-Flash model on Friday and published the weights the same day under an MIT licence, saying the gain in agent performance came from post-training and left the activated parameter count unchanged.
The release supersedes the preview version issued in April. It carries the same 13 billion activated parameters and one-million-token context window, and adds a speculative decoding module, giving it the same structure as the DSpark variant. The checkpoint is roughly 167 gigabytes. The weights are available on Hugging Face and ModelScope, alongside a technical report.
DeepSeek reported scores of 82.7 on Terminal Bench 2.1, 54.4 on DeepSWE and 68.7 on DSBench-FullStack, with several agent results exceeding those of V4-Pro-Preview, a model with 1.6 trillion parameters. The figures have not been independently verified, and the coding-agent evaluations used an internal harness that has not been released.
Because the memory requirements are broadly unchanged, institutions already running V4-Flash on their own servers can adopt the improvement without expanding their clusters. Capability gains achieved by enlarging models typically require operators to add hardware.
Several governments outside China have begun building inference capacity around Chinese open-weight models. Pakistan’s National Telecommunication Corporation signed a three-year contract in July with Data Vault Pakistan to deploy NVIDIA H200, B200 and A100 platforms in government-managed data centres, offering GPU-as-a-Service and Models-as-a-Service to federal institutions while keeping data under Pakistani jurisdiction.
Pricing was unchanged at $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit and $0.28 on output. The model natively supports the Responses API and has been adapted for Codex-style coding agents.
Most third-party hosts were still serving the April version at the time of writing, with the upgrade available through DeepSeek’s own API.
DeepSeek said the official version of its larger V4-Pro model would follow.
