Decoupled Attention from Weights - Gemma 4 26B

<!-- SC_OFF --><div class="md"><p><em>Scaling language models to handle longer input sequences typically necessitates large key-value (KV) caches, resulting in substantial memory

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1srhtvg/where_we_are_in_a_year_everything_has_changed/"> <img alt="Where we are. In a year, everything has

<!-- SC_OFF --><div class="md"><p>Google DeepMind dropped Gemma 4 today:</p> <p><strong>Gemma 4 31B:</strong> dense, 256K context, redesigned architecture targeting efficiency and

<p>Today Inception Labs dropped <strong>Mercury 3</strong>: a diffusion-based LLM with a 1M-token context window doing <strong>1,068 tokens/sec on NVIDIA GPUs</strong>, claiming

<p>The quadratic attention problem has quietly shaped everything you've built with LLMs. RAG pipelines, agentic decomposition, hybrid architectures — these aren't the natural

<!-- SC_OFF --><div class="md"><p><a href="https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF">https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF</a></p> <p><a

arXiv:2603.22300v1 Announce Type: cross Abstract: Scaling Transformers to ultra-long contexts is bottlenecked by the $O(n^2 d)$ cost of self-attention. Existing methods reduce

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1sarlb8/gemma_4_running_on_raspberry_pi5/"> <img alt="Gemma 4 running on Raspberry Pi5"