DeepSeek-V4-Pro-0813 is DeepSeek’s official release of its V4 Pro language model, replacing the preview version. It retains the preview model’s structure and adds a DSpark speculative-decoding module, with the stated goal of improving agent capabilities and production performance. The model is designed for text generation and conversational use, including tasks that involve coding, operating tools, and completing multi-step workflows. Its model card provides setup instructions for Transformers, vLLM, and SGLang, as well as access through hosted inference providers. The weights are distributed under the MIT license, though the roughly 893 GB of files make local deployment a substantial undertaking.
The reported benchmarks show large gains over V4 Pro Preview on several agent and software tasks: Terminal Bench 2.1 rises from 72.1 to 87.9, DeepSWE from 12.8 to 62.7, and Toolathlon-Verified from 55.9 to 74.1. It also scores 61.5 on NL2Repo and 83.3 on Cybergym. Results place it near or above some listed competing models on particular evaluations, but below others: for example, its Terminal Bench score is close to Kimi K3’s 88.3, while its DeepSWE score trails Kimi K3’s 67.5. The card presents the release as broadly competitive with leading proprietary systems, with especially notable improvements in coding and agent workflows.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
A Hacker News poster asks which practical AI skills individuals can learn online to benefit economically from LLMs and agents. They want specific, hands-on recommendations suited to someone with a moderately technical background, without requiring a degree. (news.ycombinator.com)
Former OpenAI researcher Will Depue claims hidden initialization patterns in six Chinese open-weight AI models match Nvidia Blackwell hardware, including systems restricted from export to China. He says the matches were reproduced on real chips and independently checked by researchers, though the claims are based on his forensic analysis. (twitter.com)
Antithesis built a system that beats NetHack on nearly 80% of game seeds within five minutes, across character builds. Successful runs take a median of under two minutes, advancing the company’s goal of creating smarter fuzzers and autonomous systems. (antithesis.com)
OpenCyclone is a free, open-source Android agent that uses your existing apps to carry out tasks through a hidden screen. It asks for approval before actions like sending messages and lets users connect their own AI models. (opencyclone.vercel.app)
Sam Bankman-Fried’s $500 million investment gave Alameda nearly 14% of Anthropic, a stake that would be worth about $37 billion at the AI company’s expected $1.5 trillion IPO valuation. The stake was sold in bankruptcy to repay FTX customers and creditors after Bankman-Fried’s businesses collapsed. (ft.com)
llama.cpp and ggml have reached v1.0, establishing stable public C APIs and clear versioning rules for breaking changes. The release strengthens the foundation for running local, multimodal AI across a wide range of hardware. (huggingface.co)
Today's best Hacker News stories, summarized and screenshotted, one email a day.