DeepSeek GSQ-RCO Coder at 1-Bit Locally
Clip title: DeepSeek GSQ-RCO Coder at 1-Bit Locally Author / channel: Fahd Mirza URL: https://www.youtube.com/watch?v=x8kK1UN5L-8
Summary
In the video, Fahd Mirza introduces the DeepSeek V4.1-Flash-1017 coder model, which contains 763 billion parameters compressed down to one bit per parameter. This compression allows the model to run locally on a single GPU consuming around 95GB of VRAM.
Mirza explains how the compression technique works by removing half of the experts that aren’t necessary for coding and agents, then using a file format called GGUF. He demonstrates running the model locally with llama.cpp and testing it on a full-stack restaurant app (Wildfare) where it fixes a bug in the food cost calculations. He also tests the model’s coding capabilities by asking it to generate a 3D-looking Döner Kebab object with fire and lighting using HTML canvas.
Overall, the video shows that despite significant compression to one bit per parameter, the DeepSeek V4.1-Flash-1017 coder model maintains impressive performance for coding tasks and local deployment while consuming relatively low resources.
Video Description & Links
Description
Running DeepSeek V4.1 Flash GSQ-RCO Coder, a 763B model pruned to about 1 bit, locally with llama.cpp and testing it on real code.
gsq rco daslab deepseek deepseekcoder
▶ LinkedIn: / fahdmirza
▶ YouTube: / @fahdmirza
0:00 What is DeepSeek V4.1 Flash GSQ-RCO Coder 1:12 bit DeepSeek Coder running locally 1:48 The Wildfare app and the hidden bug 3:10 Quantization and the size budget explained 5:15 Hermes agent tries to fix the bug 6:05 Döner kebab animation test 7:27 Nested SQL bug test 8:33 Verdict
▶ https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF
All rights reserved © Fahd Mirza