PSA: Upscaling Gemma 4 requires a proportional layer_scalar adjustment
A lot of people seem to be confused or mystified about this so figured I'd spell it out. I played around with RYS and realized that…
Latest
Gemma4 with audio input: 16.8 tok/s on Macbook M2 Max 64GB
Here's the setup I decided on for embedding gemma4-12b into a Tauri2 desktop app: Native Rust FFI into llama.cpp via llama-cpp-2 (Metal enabled) Model:gemma-4-12b-it-Q5_K_S quantized by…
First time I have seen this: my model seemed aware of its context usage ask me for compaction!
I was in a middle of a Claude Code session with GLM 5.2. Context usage 537k/1M. After finishing a task, GLM asked me this: Context note:…
[Paper] GEAR: Guided End-to-End AutoRegression for Image Synthesis
Visual generative models are typically trained in two stages. A tokenizer is first trained for reconstruction and then frozen, after which a generator is trained on…
[Paper] Multi-Block Diffusion Language Models
Block Diffusion Language Models (BD-LMs) improve diffusion-based text generation with KV caching and flexible-length generation. A natural next step is to extend them from Single-Block Diffusion…
Doing the actual math on a $20k local AI rig breakeven
Everyone under the sun says "it's free after you buy the hardware" and skips the electricity bill. Ran the numbers against a mid tier subscription to…
Is dSpark, dflash, MTP, QAT, and similar tech going to increase inference speed enough to where model spillover to disk will be more tolerable?
We’re seeing all these performance boosts coming to inference lately with things like dSpark, dllash, MTP, etc. and I know the whole model spillover-to-disk has always…
RTX5090, gemma-4-31B-it-Q6_K.gguf. Context: before – 35k, after – 80k!
Yesterday there was a message that you can increase the context for Deepseek Flash. But it turned out that everything works for Gemma4 too! function dockergemma…
google/tabfm-1.0.0
TabFM is a zero-shot tabular foundation model from Google Research. It supports classification and regression on structured/tabular data with mixed numerical and categorical columns, requiring no…
A fully local, self-hosted repo index for coding agents (Rust, MIT, runs offline)
If you run local models for coding, feeding them repo context without blowing the context window is half the battle. I built basemind to index a…