AMD's Strix Halo Successor, the 'Gorgon Halo 495 Max', May Support 192GB of Memory
According to a VideoCardz leak, AMD's next-generation APU, the Ryzen AI Max Pro 495 (codenamed Gorgon Halo), may support 192GB of memory. The current Strix Halo (Ryzen AI Max 395) offers up to 128GB of unified memory and has been a favorite in the local LLM community; if the successor bumps that to 192GB, it could run 122B-class large models with Q8 quantization at full context.
On Reddit's LocalLLaMA community, some users are already planning to run giant MoE models across current Strix Halo machines (128GB × 2 = 320GB total). If a 192GB version materializes, a single machine could comfortably run even more models, significantly expanding the options for local AI setups.
That said, this is still leak-stage information, and the final specifications may change.
Qwen3.6-35B-A3B Runs at Usable Speeds on a 5-Year-Old 6GB VRAM Laptop
A report on Reddit's LocalLLaMA community is drawing attention: a user ran Qwen3.6-35B-A3B on a 2020 ASUS ROG Zephyrus G14 (RTX 2060 Max-Q with 6GB VRAM, 24GB RAM) and achieved a generation speed of about 23 tokens per second.
The poster used llama-server (based on llama.cpp), combining a wide range of optimizations: a Q8_0 KV cache, ngram speculative decoding, NUMA isolation, context checkpoints, and more. They also shared a 128K context configuration built on Tom's fork, noting that it works even with long-context workloads.
A 35B-parameter model running at practical speeds on a five-year-old mid-range GPU is a striking illustration of just how far open-model optimization has come.
torch-nvenc-compress: Repurposing GPU Video Encoders for Tensor Compression
A library called torch-nvenc-compress has been released on Reddit's MachineLearning community. It uses the NVENC/NVDEC hardware (dedicated video encode/decode silicon) found on NVIDIA GPUs to compress AI model activations and KV caches in real time.
With NVIDIA dropping NVLink on the RTX 4090/5090, multi-GPU setups have been bottlenecked by PCIe bandwidth of roughly 30GB/s. The library's approach is to use the otherwise-idle NVENC silicon to compress and decompress data, dramatically boosting effective bandwidth.
Measured on an RTX 5090, it achieves 6.1x lossless compression on diffusion model intermediate representations and 2.7x lossless compression on LLM KV caches. A key trick is PCA-based preprocessing that transforms tensor structures into a more compression-friendly form. Running NVENC encoding and CUDA computation in parallel streams, the author measured 67% of the theoretical maximum overlap.
This promises major benefits for split-model inference in multi-GPU setups as well as distributed inference over networks. It is released under the Apache 2.0 license.
An OpenVINO Implementation of Qwen3-TTS Brings Speech Synthesis to Intel GPUs
An open-source implementation that converts Qwen3-TTS (Qwen's speech synthesis model) to OpenVINO format and runs it on Intel Arc GPUs has been released on Reddit's LocalLLaMA community.
The developer says they iterated exhaustively on data-flow analysis and device placement optimization while converting PyTorch nn.Module to OpenVINO IR format. In particular, they found the official OpenVINO documentation insufficient for handling stateful KV caches and predicting kernel fusion, making the right decisions difficult even with AI tools (including Opus 4.5).
Currently only the 1.7B model is supported, verified to run on both CPU and GPU (Intel Arc). NPU support remains future work.
Sato: A Multi-Provider AI Desktop Companion for macOS
Sato, an AI desktop assistant for macOS, was featured on Hacker News. It supports multiple AI providers and is designed to make an AI assistant easily accessible right on your desktop.
Many details about its features have yet to be disclosed, but it's worth watching as a tool that could make everyday local AI use effortless.