gemma-4

Tested the updated Gemma 4 locally using llama.cpp on an M5 Pro, achieving 60 tokens/s for coding tasks with OpenCode; good for backend but poor for UI/UX.

Google's Gemma 4 model has been enhanced to recognize when it is incorrect, improving its calibration and reliability.

Cactus Hybrid is a post-trained Gemma 4 model that outputs confidence scores, allowing on-device inference with routing to larger models when confidence is low, achieving performance comparable to Gemini 3.1 Flash-Lite with minimal calls to the larger model.

DogukanUrker demonstrates running Gemma 4 12B with full 262,144 context at ~100 tok/s on a single RTX 3060 using speculative decoding and KV cache splitting, achieving nearly full GPU utilization without CPU offload.

Google Gemma announces that developers can now use the Gemma 4 31B model as the brain for voice AI, enabled by Hugging Face and Cerebras for ultra-fast inference, as part of an open-source cascaded speech-to-speech stack.

User reports that Gemma 4 is lazy and poor at multi-turn agentic tasks compared to other models like Qwen, DeepSeek, and GPT-OSS, despite being a good chatbot.

An analysis shows that Gemma 4 models' benchmark performance is heavily influenced by chat templates rather than model weights, with template changes causing behavioral shifts without altering any parameters; notably, all sizes fail a crisis-signal scenario.

MiaAI Lab publishes a recipe to serve Google's Gemma 4 31B IT in NVFP4 quantization via vLLM, enabling 256k context, MTP speculative decoding, agentic reasoning, native tool calling, and image/video support.

Google updated Gemma 4's chat templates with major fixes to tool calling, reduced laziness, enabled Flash Attention 4 on Hopper GPUs, and released an interactive vision guide. The updates are available on Hugging Face.

Google Gemma is rolling out significant improvements to Gemma 4, driven by community feedback and contributions, as detailed in a thread.

A fix splitting a fused operation into two matmuls allows Gemma 4 to run on a 13-year-old dual-Xeon CPU at 5.2 tok/s, using only 4B of 26B weights per token.

A developer successfully runs Google's Gemma 4 26B mixture-of-experts model at about 5 tokens per second on a 13-year-old dual Xeon server without a GPU, using a modified version of ik_llama.cpp that works without AVX2 instructions.

Google announced Gemma 4 E2B optimized for the Pixel 10's TPU, enabling on-device multimodal AI capabilities like offline chat, image recognition, and audio transcription.

A variant of Gemma-4-31B uses interpretability-based steering to resist hallucination by challenging false premises, achieving double the pushback on false premises with minimal impact on benchmark performance.

An experiment demonstrating autonomous NPCs in the browser powered by Gemma 4 and E2B.

Gemma 4 can now run fully offline on-device using React Native, with local hardware acceleration via Vulkan on Android and MLX on Apple Silicon, enabling vision and tool-use capabilities like reading a flyer and scheduling events entirely on-device.

Commentary on Google's Gemma 4 technical report, noting aggressive optimization via Quantization-Aware Training and Multi-Token Prediction for faster inference.

Gemma 4 introduces a new generation of open-weight, natively multimodal language models with dense and Mixture-of-Experts architectures, featuring thinking mode for advanced reasoning, improved efficiency, and long-context capabilities.

Locagent is a private, open-source AI that runs entirely in your browser using the Gemma 4 model, ensuring conversations and documents never leave your device.

Running the quantized Gemma-4-31B model on an RTX 5090 increases context length from 35k to 80k, showcasing significant performance improvement.