Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Official5
developers.googleblog.com2025-08-14officialauthorityHN front pageHN 824 pts · 317 comments1,027 words
Google released Gemma 3 270M, a compact model with 270 million parameters designed for task-specific fine-tuning. The model features a vocabulary of 256k tokens and achieved 0.75% battery usage on a Pixel 9 Pro SoC during INT4-quantized testing. It provides production-ready quantization for INT4 precision with minimal performance degradation.written, reviewed, cited
- The model includes 170 million embedding parameters due to a large vocabulary size and 100 million for transformer blocks.
- An instruction-tuned model is released alongside a pre-trained checkpoint for developers needing different training starting points.
developers.googleblog.com2025-04-18officialauthorityHN front pageabout the gemma 3 familyHN 602 pts · 276 comments1,077 words
Google released Gemma 3 Quantization-Aware Training (QAT) models that incorporate quantization during training to reduce memory usage while maintaining accuracy. Applying QAT on ~5,000 steps reduces the perplexity drop by 54% when quantizing down to Q4_0, allowing the 27B variant to fit in 14.1 GB of VRAM instead of 54 GB. These int4 models run efficiently on consumer hardware like the NVIDIA RTX 3090 and are available via Ollama, llama.cpp, and MLX.written, reviewed, cited
- Deploying the largest models often requires lower-precision formats like FP8, potentially accepting a performance trade-off for feasibility.
- The Gemmaverse offers many alternatives using Post-Training Quantization (PTQ) with contributions from members such as Bartowski, Unsloth, and GGML.
developers.googleblog.com2025-06-26officialauthorityHN front pageabout the gemma 3 familyHN 405 pts · 191 comments1,690 words
Gemma 3n is a mobile-first AI model family designed for on-device use with multimodal capabilities. The E2B and E4B versions have raw parameters of 5B and 8B but operate with as little as 2GB and 3GB of memory respectively. It supports 140 languages for text and includes a MobileNet-V5-300M vision encoder that delivers a 13x speedup with quantization, requires 46% fewer parameters, and has a 4x smaller memory footprint.written, reviewed, cited
- Developers can create custom-sized models between E2B and E4B by adjusting feed forward network hidden dimensions from 8192 to 16384.
- The Gemma 3n audio encoder processes clips up to 30 seconds at launch but supports arbitrarily long audios with additional training.
developers.googleblog.com2025-05-20officialauthorityHN front pageabout the gemma 3 familyHN 441 pts · 163 comments904 words
ai.google.dev2025-03-23officialauthorityHN front pageabout the gemma 3 familyHN 202 pts · 32 comments2,144 words
Analysis5
github.com2026-07-04authorityabout the gemma 3 familyHN 6 pts · 2 comments193 words
MetalChat is a Metal-accelerated C++ framework and command line interpreter for inference of Meta Llama, Google Gemma, and Liquid models. The library API and CLI are under active development, therefore they may change without any deprecation notice. The framework is distributed under GPLv3 license and can be installed using Homebrew package manager with brew tap ybubnov/metalchat.written, reviewed, cited
unsloth.ai2025-03-19authorityHN front pageabout the gemma 3 familyHN 226 pts · 78 comments
storage.googleapis.com2025-03-12HN front pageabout the gemma 3 familyHN 488 pts · 245 comments
github.com2025-08-20authorityHN front pageHN 417 pts · 57 comments327 words
github.com2026-01-26authorityabout the gemma 3 familyHN 65 pts · 23 comments2,207 words
