Most people building local AI focus on Nvidia, but AMD has great stuff too! Here are some people to follow:


@dcapitella: The GOAT. His Strix Halo containers let you keep the host system clean and instantly switch between different llama.cpp forks and backends. Also has a YouTube channel.

Strix Halo toolboxes | YouTube | Benchmarks



@Italianclownz: Created AMD-native quant format ROCmFPX. If you aren't using it, you're leaving a lot of performance on the table compared to stock llama.cpp. Almost everyone on this list uses his work on ROCmFPX.

ROCmFPX | HuggingFace | YouTube



@ciruai: Creates and optimizes ROCmFPX quants, benchmarks focusing on agentic workflows

Inference lab | HuggingFace



@laurent_zw: Creates and optimizes ROCmFPX quants, maintains custom ROCmFPX fork

HuggingFace | ROCmFPX fork | Agention AI Lab



@zetaphor (BlueSky): His blog is a goldmine of experiments and how-tos with Strix Halo. Updates regularly.

Blog | GitHub



@pupposandro: creating @luceboxai. Trying (and succeeding) at building high-perf hybrid (Strix Halo + discrete GPU) AMD setups

Lucebox



@1337hero: Benchmarks and optimizations for AMD GPUs plus ROCmFPX quants. May or may not have contributed to me choosing AMD

HuggingFace | Website



@rc_morano: Creates and optimizes ROCmFPX quants

HuggingFace



@NathanW1014: llama.cpp fork tuned for long-context, quantized-KV workloads, considerably better Vulkan performance on Strix Halo.

llama.cpp fork | Strix Halo llama.cpp toolbox



@1bitlabs: Building native ternary (1-bit) models from scratch so dense 100B-class models can run locally. Trained experimental Ultron 0.3B base model

HuggingFace | Ultron 0.3B



@schuttdev: Building hipfire, a Rust + HIP inference engine for consumer RDNA and Strix Halo that skips the full ROCm library stack

hipfire repo | GitHub | HuggingFace



@high52weeks: Writes focused, single-GPU, single-model kernels for Strix Halo that beat the usual llama.cpp forks

halogen-flash-server (Qwen 3.8 Flash) | halogen-server (Qwen 3.8 27b) | HuggingFace



@Plunderstruck: Creates and optimizes ROCmFPX quants

HuggingFace



KingJones (not on X): ROCmFPX and NVFP4 quants for Strix Halo. Runs and publishes benchmarks including results that didn't work

HuggingFace | Benchmarks