DeepSeek V4 GGUF quietly drops on Hugging Face via antirez
Antirez slipped DeepSeek-V4-Flash weights into the old folder while users waited.
Deep Dive
Antirez stealthily uploaded new weights into the old folder—and everyone was left tapping their fingers, waiting. The files landed in the Hugging Face repo at https://huggingface.co/antirez/deepseek-v4-gguf/tree/main, shared by /u/challis88ocarina.
Key Points
- Antirez uploaded DeepSeek-V4-Flash GGUF weights quietly to the existing deepseek-v4-gguf Hugging Face repo.
- The file mixes Q4K experts with F16 compressor/indexer and Q8 attention/shared/output layers for quality.
- IMatrix calibration (imatrix-0731) improves quantization accuracy, making local inference more viable.
Why It Matters
Local execution of DeepSeek V4 on consumer hardware is now possible without waiting for official release.