BeeLlama.cpp v0.4.1: KVarN, KV precision tail, q2_0-q3_1 KV cache, improved supp

<!-- SC_OFF --><div class="md"><p>I applied video compression to LLM inference and got <strong>10,000x less quantization error at the same storage cost</strong></p> <p><a

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s5cggh/google_turboquant_blew_up_for_kv_cache_heres/"> <img alt="Google TurboQuant blew up for KV cache.

<blockquote> <p><em>Originally published at <a href="https://llmkube.com/blog/turboquant-m5-max-quality-and-asymmetric" rel="noopener

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s92x7z/attnrot_ggerganovs_turboquant_lite_is_on_the_cusp/"> <img alt="attn-rot (ggerganov's

<!-- SC_OFF --><div class="md"><p>I tested two llama.cpp builds on the same <strong>Qwen3.6-27B-NVFP4</strong> model.</p> <p><code>llama-bench</code> reports the model label as

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1sl59qq/updated_qwen359b_quantization_comparison/"> <img alt="Updated Qwen3.5-9B Quantization Comparison"

<!-- SC_OFF --><div class="md"><p>Last year researchers affiliated with NVIDIA, University of Warsaw, and University of Edinburgh published <a

<!-- SC_OFF --><div class="md"><p>Every output token the model generates requires reading the entire KV cache from High Bandwidth Memory.</p> <p>Here's what that actually looks