Search Articles — Sudonull

Search Results

In this project

Gemma 4: open MoE models from Google up to 31B parameters

https://sudonull.com/gemma-4-open-moe-models-from-google-up-to-31b-parameters

Explore Gemma 4 from DeepMind: MoE architecture, multimodality, edge deployment. 31B outperforms large models. Download weights, integrate into projects for developers.

Quantization Gemma 4 31B under 57GB Kaggle

https://sudonull.com/quantization-gemma-4-31b-under-57gb-kaggle

Learn how to load 62GB Gemma 4 31B model on Kaggle with 57GB limit: NF4, bitsandbytes and memory hacks. Get 18GB artifact for local inference. Instructions for developers.

Fine-tuning Gemma 4 on Cloud Run with GPU

https://sudonull.com/fine-tuning-gemma-4-on-cloud-run-with-gpu

Instructions for fine-tuning Gemma 4 31B on Cloud Run Jobs for image classification. QLoRA, LoRA rank 64, accuracy 94%. For middle/senior developers. Start fine-tuning now.

Qwen 3.6 vs Gemma 4: testing on a local machine

https://sudonull.com/qwen-3-6-vs-gemma-4-testing-on-a-local-machine

Performance comparison of Qwen 3.6 and Gemma 4 on a laptop with RTX 4070. LM Studio setup, VRAM optimization, test results. Read the guide.

llama.cpp: ASR in Gemma 4 and Qwen3

https://sudonull.com/llama-cpp-asr-in-gemma-4-and-qwen3

llama.cpp integrated speech recognition Gemma 4 (E4B/E2B) and Qwen3. Launch on GPU, prompts, parameters for stability. Instructions for middle/senior dev.

Trending Now