DeepSeek just open-sourced something that isn't a model.
An AI model is a brain in a jar. Smart, but no hands, no eyes, no memory. A harness is the body around it. DeepSeek built that body entirely out of LEGO. Every piece is a plugin: tools, memory, sandbox, UI, even the loop
Sakamakismile/Qwen3.8-27B-MTP-NVFP4
Thank you for publishing the model and quantization model. SGLang/Docker/Opencode Complete the race with 1 prompt like vLLM 1h31min make 3d voxel game using javascript save as q38-27b.html The speed of DeepSeek is a must-have. recipe
Execution result of "Neo Edo City Prompt (named)" on DeepSeek V4 Flash 0731 Max.
It's very well made, but compared to Pagoda Garden, it seems to have a lot more fluctuation. But hey, it might be better than Fugaku Sanjurokkei.
The wildest jump was DeepSWE.
Previous score: 7.3 New V4 Flash: 54.4 Same architecture. Same parameter count. DeepSeek retrained the model and massively improved its agent performance.
You have reached the end of the archive
All of deepseek