Has extended the coding session limit of deepseek-v4.1-flash to 5M tokens, achieved by inference-time context compression.
At Max thinking effort, achieves 70%+ accuracy on DeepSWE. - at Low thinking effort, significantly outperforms GLM-5.2 and uses 30% less
I really abuse AI too much 🤣 deepseek it delights me
DeepSeek-V4.1-Flash is handling soo many different tasks without a sweat.
I really love this one, a Three.js pastel themed village, with flying people. Done in one-shot, result is really good, high quality colors, added rain and sunset theme made this x10 better. Via DeepSeek
I spent the weekend getting a 552B model to run across three DGX Sparks cabled to each other in a ring, no switch in the room.
It works, it decodes at parity with the reference numbers, and the whole thing is one declarative object on Kubernetes now. LLMKube 0.9.27 is out with
You have reached the end of the archive
All of deepseek