IT之家 9 月 26 日消息,OpenCode 昨日宣布启动“Operation Cheepseek”第二阶段,DeepSeek V4.1 Flash 在 OpenCode Go 中原本限时给出的 60 美元(IT之家注:现汇率约合 403.5 元人民币)额度现调整为永久有效。
此前,OpenCode 曾将 DeepSeek V4.1 Flash 的使用额度临时提高至原标准的 4 倍,即从 15 美元提高至 60 美元,并计划于 9 月 20 日结束该活动。OpenCode 当前 Go 页面已将该模型的额度标注为 60 美元,原 15 美元额度被划线标记。
DeepSeek 最新发布的 V4.1 Flash 使用 5520 亿参数的 MoE 架构,输入侧激活参数量为 80 亿,输出侧为 160 亿,并使用新的 Causal Encoder-Decoder 架构。模型同时兼容原生多模态视觉理解。
与上一代相比,DeepSeek V4.1 Flash 的 KV Cache 对 HBM 的需要降低至四分之一,对 SSD 的需要降低至八分之一。
DeepSeek 此前还宣布 OpenCode 已作为官方合作伙伴全量接入 V4.1 Flash。OpenCode 数据表明,截至 9 月 25 日,该模型在平台近期使用量中排名第一,占观测到的 200 万级使用量的约 13%。
本文来源:https://www.ithome.com/1/007/421.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT Home, September 26 - OpenCode yesterday announced the launch of the second phase of "Operation Cheepseek". The $60 credit (IT Home note: at the current exchange rate, approximately 403.5 RMB) originally offered on a limited-time basis for DeepSeek V4.1 Flash in OpenCode Go is now adjusted to be permanently valid.
Previously, OpenCode temporarily increased the usage credit for DeepSeek V4.1 Flash to 4 times the original standard, from $15 to $60, and planned to end the promotion on September 20. OpenCode's current Go page now lists the model's credit as $60, with the original $15 credit crossed out.
DeepSeek's latest V4.1 Flash uses a MoE architecture with 552 billion parameters, with 8 billion activated parameters on the input side and 16 billion on the output side, and employs a new Causal Encoder-Decoder architecture. The model is also compatible with native multimodal visual understanding.
Compared to the previous generation, DeepSeek V4.1 Flash's KV Cache reduces the need for HBM to one quarter and the need for SSD to one eighth.
DeepSeek also previously announced that OpenCode has fully integrated V4.1 Flash as an official partner. OpenCode data shows that as of September 25, the model ranks first in recent platform usage, accounting for approximately 13% of the observed 2 million-level usage.
