Gemma 3

Developer-heavy sample (Hacker News) running Gemma 3 locally across the size range. Strong positives cluster around the 27B instruct model: one user calls gemma-3-27b-it-qat-4bit their 'new favorite local model' alongside Mistral Small 3.1 24B, running on an M2 64GB via Ollama (~22GB) or MLX (~15GB) with memory to spare; another finds the Q4_K_M 27B responses better than DeepSeek-R1-Distill-Qwen-32B and its image recognition 'stronger than I expected.' A Windows/ROCm user reports ~26 tokens/sec at FP16 on a 16GB Radeon 7900 GRE and says 'I just love it,' while still preferring QwQ 32B for coding (unusable at ~3 tok/s on that card). Multilingual strength is noted: 'one of the few models that can write Norwegian decently' with good instruction following. Practical multimodal use exists: one user has Gemma describe photos and emit keyword lists in production, needing only prompt tweaks to strip preamble like 'Here's a description of the image.' The tiny 270M variant draws the harshest criticism: it answers 'Mount Everest' to the tallest, second, third AND fourth tallest mountain questions, fails simple animal-listing and alphabetizing tasks ('Out of these characteristics I can observe only speed'), though some defend it as a deliberately specialized sub-1B model and others enjoy it as a fast 241MB toy. Censorship is a recurring complaint: a game operator reports Gemma mishandling player messages like 'I will kill you', hoping it's fixed in a hypothetical Gemma 4; 'LLM abliteration' workarounds are discussed openly. Tooling friction appears: Ollama reportedly doesn't quantize the embedding table (kept at FP16), pushing Mac users toward llama.cpp, LM Studio or MLX. A speed-vs-cloud debate runs through threads: cloud APIs answer faster than reading speed, so local runs are justified mainly by privacy and training-data-consent concerns. Notably, Google engineers participate in the threads, explaining the dense architecture was chosen for 'simplicity and finetunability' and that community adaptability drove embedding-size decisions.

No aggregate ratings were found for this product during the last harvest.