event
DeepSeek Release
-
Reusing key--value (KV) caches across language model updates could avoid repeated computation of shared context, but changes in model weights raise a fundamental question about what these caches preserve. We investigate direct cache reuse…
-
We shrank DSV4-Pro by 24%, within 0.5pt of original on MMLU – Tileward (tileward.com via hn) mmludeepseek 6d
Our TWX3 build of DeepSeek-V4-Pro is 24% smaller than the model as published, within half a point of it on MMLU, and level on tool calls and made-up facts.
-
Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU (github.com via hn) moellamadeepseek 8d
♟️🐋 llama-modchip (v2.4.0) Break the 64 GB RAM Wall: Run 284B DeepSeek-V4-Flash ( 284.33B Frontier MoE params 56.56 GiB checkpoint on disk 57.61 GiB peak VmRSS @ 256K ctx 44.36 tok/s 284B CPU prefill 60.4–111.6 tok/s 30B MoE CPU decode 13.…