448 HN pointsWhy trending
Relative attention score: 40.9/100. Recent activity and cross-channel evidence affect position. This is not a quality rating.
- 448 HN pointsHacker News · observed 9 Nov, 08:15 pm IST
- 196 commentsHacker News · observed 9 Nov, 08:15 pm IST
How ranking works · Popularity can reflect hype, controversy or manipulation.
01 / The short version
What happened
llama-modchip v2.4.0 enables CPU-only inference of large models without GPUs. The publisher claims DeepSeek-V4-Flash 284B runs on 64 GB RAM using 56.56 GiB disk space and 57.61 GiB peak VmRSS at 256K context, with prefill speeds of 44.36 tok/s and decode speeds of 9.06–14.7 tok/s. Nemotron-3.5-30B-A3B reportedly runs at 25.0–60.4 tok/s on 16-core Zen 5 hardware using 14.35 GiB RAM. The software uses Fractal-64 quantization and Stockfish-PVS speculative decoding. A pre-compiled Linux binary and Hugging Face GGUF models are available via GitHub Release v2.4.0. Performance claims are attributed to the publisher and have not been independently validated.
See the exact references02 / Key takeaways
What you need to know
- 01
DeepSeek-V4-Flash 284B reportedly fits in 64 GB RAM at 57.61 GiB VmRSS.
- 02
Nemotron-3.5-30B-A3B reportedly achieves 25.0–60.4 tok/s on pure CPU.
- 03
Custom GGUF tensor type 40 requires patched binaries and will fail in unpatched llama.cpp, Ollama, or LM Studio.
Keep in perspective
What to watch for
Performance figures are publisher-claimed; independent benchmarks are needed to verify speed and memory usage.
Go to the source
Exact references
These are the original pages used for this brief. Publisher claims are not independent evaluations.
01Primary source · github.comOriginal linked resourcehttps://github.com/synchopate/llama-modchip02Reporting · Hacker NewsOriginal Hacker News discussion and engagementhttps://news.ycombinator.com/item?id=5041196103Documentation · GitHubRepository README used for this summaryhttps://github.com/synchopate/llama-modchip/blob/main/README.mdAI-generated from the linked source. It can miss context; verify consequential details in the original. How the radar works