GPU
5 updates on GPU.
Google adds local model support to the Antigravity SDK with Gemma 4 on LiteRT
Antigravity agents can now run on Gemma 4 26B A4B through LiteRT, with Google recommending more than 24 GB of VRAM or unified memory.
Perplexity brings its local Portable Computer agent to AMD Ryzen AI Max PCs
Portable Computer runs Qwen 3.8 27B or PPLX 27B locally on Windows PCs with AMD Ryzen AI Max chips and at least 24 GB of GPU-accessible memory.
Hugging Face adds native GGUF inference to Transformers, starting with Qwen3.5 on Macs
Transformers now loads GGUF models in packed form and runs them on Apple Silicon GPUs via Metal kernels, starting with Qwen3.5.
S2-MoE speeds up MoE decoding by up to 5.3x in llama.cpp
S2-MoE is a self-speculative decoding method for mixture-of-experts models that reaches up to 5.3x speedup and about 2.0x on average.
MobiBench paper benchmarks small LLMs in llama.cpp on two laptops and an RTX 3060
MobiBench is a llama.cpp benchmark suite for speed, memory and accuracy of small LLMs, but the reported tests ran on two laptops and not on phones.