Strix Halo is AMD's processor platform for local AI inference, with the Ryzen AI Max+ 395 as a representative part. Recent coverage focuses on running local models on it: halogen 0.12.0 raised decode at 1M tokens of context from 27.3 to 38.3 tok/s on a 128GB Ryzen AI Max+ 395, Qwen3.8-Flash-Next gained official Strix Halo support on Strata with up to 1M context, and Sipeed announced MaixStation as comparable to Strix Halo and Nvidia's DGX Spark.
Sipeed announces MaixStation, comparable to Strix Halo and DGX SparkSipeed has announced MaixStation, positioned as comparable to AMD's Strix Halo and Nvidia's DGX Spark. The news comes via Reddit's LocalLLaMA community, with specifications and availability not yet disclosed.
LemonSeed: Qwen3.8-27B Hits 159 tok/s on iPad via eGPU R9700LemonSeed Studio is an iPad editor that runs on-device inference via an AMD GPU in a Thunderbolt enclosure, embedding the upstream Linux amdgpu driver. Running Qwen3.8-27B Q4 with a DFlash2 draft, the R9700 decodes at 158.5 tok/s on iPadOS and 159.4 tok/s on macOS, versus a 31.2 tok/s baseline, while Strix Halo reaches 64.4 tok/s on Linux.
Qwen3.8-Flash-Next gains Strix Halo support on StrataA developer released official Strix Halo support for Qwen3.8-Flash-Next on Strata, claiming best long-context decode performance, up to 1M context length, Linux-only and marked experimental.
Qwen3.8-Flash-Next Hits 1M Context on Strix HaloDeveloper releases halogen 0.12.0, fixing degradation at long context depth. On a Ryzen AI Max+ 395 with 128GB RAM, decode at 1M tokens of context improved from 27.3 to 38.3 tok/s, with prefill taking about 18 minutes cold.
Local LLM community glimpses a golden eraA Reddit post says hardware shortages have pushed the local LLM community back into low-level optimization: forks of llama.cpp and Strix Halo's halogen-flash-server boosted Qwen 3.8 Flash Next (Q38FN) to about 52 tok/s decode and 1300 tok/s prefill. The author argues scarcity spurs learning and technical growth.
The A.I. Daily