|
查看: 101|回复: 2
|
Deepseek 4预览版 技术含量
[复制链接]
|
|
|
Deepseek 4 Pro just dropped, and it's BEATING all the closed-door models. This is going to be a kick in the nuts for all the closed-door big boys...
This morning, I was talking about how the last 72 hours have been insane and that Deep Seek was about to drop any moment.
Literally a few hours later, boom, here it is.
Now you probably don't have the hardware to run this because it's a freaking beast. It actually BEATS with Opus 4.6 Max, GPT 5.4 X High, and even Gemini 3.1 Pro High.
1.6T total parameters with 49 billion active parameters, 1 million context window, 3.7 times lower FLOPs, and 9.5x smaller KV cache than V3.
Trained on 32 trillion plus tokens, and then there's its baby brother v4 flash. 284 billion total parameters with 13 billion activated as MOE. Same 1-million-context window, and all of this under an MIT license.
This is a kick in the nuts for the premium American big boys.
The interesting thing for me personally is that the benchmarks on live Codebench, Codeforces, and Apex shortlists all go to Deep Seek v4. First time an open-weight model has taken the crown in competitive programming. It beats the crap out of the big boys but has not been benchmarked against ChatGPT 5.5 (released yesterday) and 4.7 (released last week).
However, I'm gonna say it, who cares? Opus 4.6 and ChatGPT 5.4 Max and Gemini 3 Pro are already bloody awesome, and this comes in at a fraction of the price.
Now, the interesting thing is that I expect GLM and qwen to respond within weeks, because the Chinese open-source stack is now leading in coding. The closed labs have to justify their pricing, which is ironic given that top-end ChatGPT just doubled to $180 per million tokens, while Opus 4.7 is burning more tokens for the same $75 per million output.
Honestly, I am agnostic on all these models. If I'm getting the same power with DeepSeek that I do with Opus 4.6 and 5.4, or with new models as they come out, I'm quite happy to save 95% of my costs.
Not only that, they don't have rate limits, which means I can hammer the shit out of it with all of my projects and coding and get exactly what I need. I've been saying for weeks that Deep Seek is going to change the game, and to watch this space. I even repeated that this morning.
I just didn't expect it would be today, so that's going to keep me busy for the next few days as I put it through its paces on my own setups and force Claude Code and others into bending to my will. |
-
|
|
|
|
|
|
|
|
|
|

楼主 |
发表于 25-4-2026 08:35 AM
来自手机
|
显示全部楼层
|
|
|
|
|
|
|
|
|
|
发表于 25-4-2026 09:33 AM
|
显示全部楼层
|
但是 deepseek 大多数建议叫你使用中国网站。 |
|
|
|
|
|
|
|
|
| |
本周最热论坛帖子
|