Rare find
Deepseek.cpp. Run DeepSeek AI models on your computer without a graphics card.
github.com/andrewkchan/deepseek.cpp
Vaya's read on this project
Problem, audience, market, and the verdict — sign in to see it.
Updates
October 2025
- tested v3.1, terminus
- type union operator unsupported for python<3.10
June 2025
- add nucleus (top-p) sampling
May 2025
- quality info and known issues
- add v3 encoded wikitext
- rename
- add cli arg -w to use wikitext in perplexity mode
- test perplexity with 1 v2-lite-encoded chunk of wikitext-2
- half chunk size
- add --bsize arg to convert.py to allow configuring f8e5m2 block size
- apply attn sinks + rotation if past original_max_position_embeddings
- add rope scaling params to file format
- add tool to dump debug_map in HF safetensors format
- expose dump_debug_map
- add better err msg for perplexity mode
- (non-back-compat) remove cap on max_seq_len and read it from tokenize…
- this was accidentally deleted in a refactor
- fix bugs with f16/f32, v2-lite active_bytes
- oops
- refactor active_bytes (should also fix bugs w.r.t. dense layers)