第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
DeepSeek[3][5]的新闻发布日期为2026年4月24日,开头写道"DeepSeek-V4 Preview 正式上线并开源";其2026-04-24的API更新日志写道"DeepSeek API现已支持V4-Pro和V4-Flash",OpenRouter也在同一天上架了这两款模型。
DeepSeek's[1][3][5] model card: V4-Pro 1.6T parameters (49B activated) and V4-Flash 284B (13B), 1M-token context, MIT license, a hybrid of Compressed Sparse and Heavily Compressed Attention that needs 27% of V3.2's single-token FLOPs and 10% of its KV cache at 1M tokens, 32T+ pre-training tokens and the Pro-Max benchmark tables.
DeepSeek's[1][5] own changelog: 2026-04-24 V4 launch with deepseek-chat and deepseek-reasoner routed to V4-Flash until 2026-07-24, then the 31 July V4-Flash update, the 13 August V4-Pro GA with low/high/max thinking effort and peak/off-peak pricing from 16 August, and V4.1-Flash on 10 September.
The V4 technical report (DeepSeek[1][3][5]-AI, 319 authors): hybrid attention, manifold-constrained hyper-connections and the Muon optimizer, over 32T pre-training tokens, and the claim that V4-Pro-Max redefines the state of the art for open models.
DeepSeek's[1][3] current price table: V4-Pro-0813 at $0.66 off-peak and $1.32 peak per million cache-miss input tokens and $1.98 / $3.96 output, 1M context and 384K maximum output; shows how far the launch prices have since moved.
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
DeepSeek V4 is a very, very inexpensive model.
V4 Pro now enters as the #2 open weights reasoning model on the Artificial Analysis Intelligence Index, behind only Kimi K2.6 (54).
V4-Pro Max is 3 to 15 points behind, with the biggest gaps on the agentic-coding surfaces DeepSeek markets as V4's strength.
V4 still trails Anthropic's Opus 4.6 on MRCR 1M long-context recall (92.9%) and trails Kimi K2.6 on the newer SWE-bench Pro.
原标题: DeepSeek-V4-Pro and DeepSeek-V4-Flash Released