DeepSeek下一波大模型被指即将问世：3万亿参数 性能超GPT-6 Astra
快科技9月13日消息，本周DeepSeek发布了V4.1 Flash，这个大模型架构变化之大，完全可以称得上是V5了，但DeepSeek很低调地没换大版本号。
Pro系列的新品确定了是V4.1 Pro，具体升级还没公布，如果是V4到V4.1这样的提升幅度，那V4.1 Pro性能上限也会很高，但考虑到之前V4 Pro-0813正式版的不及预期，大家期待值不要拉太高。
后续还有一波大模型，爆料者称它叫做DeepSeek Code 2.0，即将发布，时间窗口是9月份，预计性能会击败Mythos 5.1和GPT-6 Astra，这是当前最强大的两个前沿级大模型。
其他方面，DeepSeek Code 2.0预计会继续保持开源开放，设计重点是Computer Use，也就是电脑控制，AI替代你操作电脑完成很多工作，Astra目前很多让人惊讶的使用就来自这一方面。
这个Code 2.0拥有高达3万亿以上的参数量，显然这也是它提高性能的关键，这方面国内目前最强的是K3大模型的2.8万亿，千问的Qwen 3.8 Max是2.4万亿参数量，DeepSeek V4 Pro是1.6万亿参数。
从V4.1 Flash换用新结构使得参数量翻倍来看，这个Code 2.0从当前1.6万亿提升到3万亿参数的可信度是有的，想跟Mythos 5.1及Astra这样的大模型竞争就需要这个级别的参数量，不然性能上限是很难提升的。
这个爆料的可信度不好说，DeepSeek过去的23-24年倒是推出过三款名字带Code的大模型，不过那时候表现平平，这次要是重推偏向代码/逻辑的Code大模型，作为旗舰级产品倒也不失为一个方向，就让V4.1 Pro及Flash面向通用Agent任务即可。
本文来源：https://ishare.ifeng.com/c/s/8wObTubv1rG
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
The next wave of DeepSeek models is said to be imminent: 3 trillion parameters, outperforming GPT-6 Astra.
09/13 22:52 Updated 2026-09-14 06:08:07 Official account of 快科技 From Henan
快科技 reports on September 13th that this week, DeepSeek released V4.1 Flash, which underwent such significant architectural changes that it could easily be classified as V5, but DeepSeek remained modest and did not increment the major version number.
The new product in the Pro series is confirmed to be V4.1 Pro, with specific upgrades yet to be announced. If the improvement follows the trend from V4 to V4.1, the ceiling of V4.1 Pro will be very high; however, given the lackluster performance of the previous V4 Pro-0813 final version, expectations should be tempered.
There is a subsequent wave of models coming, rumored to be named DeepSeek Code 2.0, scheduled to launch within the September window. It is anticipated to surpass Mythos 5.1 and GPT-6 Astra, currently the two most powerful frontier-level models.
Regarding other aspects, DeepSeek Code 2.0 is expected to remain open-source and open, with a design focus on “Computer Use”, meaning controlling the computer to perform tasks—a domain where Astra has already shown remarkable capabilities.
This Code 2.0 boasts a massive parameter count exceeding 3 trillion, which is clearly the key to its improved performance. In this regard, the current domestic leader is K3 with 2.8 trillion, followed by Qwen 3.8 Max at 2.4 trillion, and DeepSeek V4 Pro at 1.6 trillion.
Given that V4.1 Flash doubled its parameter count through structural changes, the possibility of Code 2.0 scaling from 1.6 trillion to 3 trillion is credible. To compete with models like Mythos 5.1 and Astra, such a magnitude of parameters is necessary; otherwise, performance ceilings cannot be raised.
The credibility of this rumor is uncertain. Over the past 23-24 years, DeepSeek has launched three models with “Code” in their names, but they were unremarkable. If this new iteration is indeed a flagship-level product focused on code/logic, it represents a valid direction. Meanwhile, V4.1 Pro and Flash can remain dedicated to general-purpose agent tasks.