Deepseek V4 flash just beat its own pro model on all 9 agent benchmarks.
And after 50+ builds, the benchmark scores weren't even the most interesting part. The numbers: → DeepSWE jumped from 7.3 → 54.4 after retraining → Terminal Bench 2.1: Flash scored 82.7 vs Pro at 72.1
Unboxing the Dell Pro Max 16 Plus powered by the NVIDIA RTX PRO 5000 Blackwell.
Running open models like DeepSeek & Kimi locally — no API, no internet — alongside robotics simulation and Physical AI workloads.
I'm running Qwen3.8 27B on Pi, but it still looks like this.
Quite late. Deepseek V4 Flash might be too practical to leave behind. If it were more optimized, it might be able to win overall, but at this point in time, Deepseek V4 Flash might be too powerful when it comes to coding.
DeepSeek V4 Pro vs Claude Fable 5 vs Grok 4.6: picking wrong costs 50x more.
3 frontier models. 1 number nobody is talking about: 276. 🤯 Agents reread your instructions hundreds of times per task. DeepSeek's cached
You have reached the end of the archive
All of deepseek