AI Briefing#gemma
The latest AI and developer news about #gemma, with the original source and a short summary.
The latest AI and developer news about #gemma, with the original source and a short summary.
Resolved Gemma 4 E2B multimodal OOM in llama.cpp on iOS.
A llama.cpp benchmark comparing ROCm 6.4.2 and Vulkan on the 6900 XT was published.
Pocket LLM v1.5.0 has been released, supporting voice and image input as well as OCR.
A single H100 80GB benchmark confirmed the advantage of Gemma 4 E2B and Qwen 3.6 A3B FP8.
In a single H100 GPU benchmark, Gemma 4 E2B-it delivered the best results in both throughput and TTFT.
Gemma was sensitive to KV cache quantization, while Qwen held up relatively well.
Built a chat UI for Gemma-4-E4B using a Qwen-series model in about 8 minutes.
Gemma4 and Qwen 3.5/3.6 were compared on RTX 5090 using the same document-writing task.
A procedure for running Gemma 4 VLA locally on a Jetson Orin Nano Super was shared.
In a 108K-token JS file memory test, Qwen 3.6 significantly outperformed Gemma 4.
Using Hadamard rotation and a Lloyd-Max codebook, the PLI of Gemma 4 E2B was compressed to 2 bits.
Google has released LiteRT-LM, its on-device LLM inference engine.
Running the same prompt across 9 local models revealed a correlation between quantization provider differences and code length.
Even 'uncensored' models have a flinch that lowers the probability of banned words on their own.
On Gemma 4 26B-A4B GGUF, Unsloth quants top the charts by KLD in most cases.
In the browser, Gemma 4 E2B converts prompts into Excalidraw diagrams.
Llama.cpp token speed measurements on the Zenbook A16 with Snapdragon X2 Elite Extreme.
In a 30k LOC agentic evaluation, Qwen 3.6 35B outperformed Gemma 4 26B.
In Open WebUI with OpenUI attached, gpt-5.4-mini responded the fastest.
Compared MMLU scores of GGUF quantized models in a 24GB VRAM environment.
A fast, uncensored local model optimized from Gemma 4 26B into MLX 4-bit.
On an RTX 5080·64GB Windows 11 setup, Llama.cpp was about 2x faster than LM Studio.
In Gemma 4 26B a4b, multi-step think/tool loops break off midway.
Ran Gemma 4:26B on a Mac Studio M4 Max using OC + LM Studio, achieving 40 tok/s.