
Links: All the data collected across 51 models: loom-fun.github.io/loom-dev Loom skills: github.com/loom-fun/loom-ski… Loom compiler and server: github.com/ROCm/hrx-system HF collection of top-51 models ported to Loom: huggingface.co/loom-fun

Local AI server for on device and on prem intelligence.

Links: All the data collected across 51 models: loom-fun.github.io/loom-dev Loom skills: github.com/loom-fun/loom-ski… Loom compiler and server: github.com/ROCm/hrx-system HF collection of top-51 models ported to Loom: huggingface.co/loom-fun

Port and optimize your own model: /loom get <hf-org>/<model-gguf> working on loom-serve, optimize it, and report the performance relative to llamacpp rocm

Fetch and serve a model: pip install "hrx-loom[hf]==0.1.0" loom-fetch huggingface.co/loom-fun/bart… --output=qwen3.8-27b loom-serve --model=qwen3.8-27b/model.json --device=amdgpu --memory_bytes=52gb

The speed is real too. On 32K-token prompts, prefill was 1.5x faster than llama.cpp ROCm at the median, up to ~2.1x, and ~1.9x faster than Vulkan. Each prompt had a unique nonce so prefix caching couldn't help.

We have had pretty good success over a broad range of mdoels - 51 models across 19 architectures, from Qwen3.8 and Gemma 4 to MoEs up to 122B and vision models. Each port is on Hugging Face with its .loom files.

The know-how ships as loom-skills, so Claude Code, Codex or Pi can write and serve optimized model programs. loom-serve swaps a model or its kernels without a restart.

The novelty is the compiler. Loom puts each model's code in small, readable .loom kernel files instead of burying it in an inference engine. Agents can write, test and tune those files directly.

hrx-loom is a new end-to-end tool stack (compiler, runtime, serving tools) that lets u build, optimize, serve, and share model programs without any modifications to the inference engine or server.

Getting a new LLM architecture to run optimally meant development and optimization of kernels for weeks. @AMD 's experimental hrx-loom cuts that down to hours. 🧵


Halogen 0.20.0 is out for Strix Halo • OvisOCR3 document parsing on the @AIatAMD NPU (your own weights) • OCRBench 85.7 vs 85.3 upstream • ROCm 10.1: ~9–15% less energy, same speed • 2–6% faster prompt reading on Qwen3.8-Flash-Next github.com/peonist-ai/haloge…

The @AMD @AIatAMD Backed @lemonade_server Lemonade Local AI Server Ships Redesigned Desktop App & Web UI phoronix.com/news/Lemonade-2…

Lemonade app got a refresh!! Try the sick new UI at: github.com/lemonade-sdk/lemo… youtu.be/mg2PqmRujMo?si=JfIr…

I want to thank the @lemonade_server team for including ROCmFPX and thank @dcapitella for making it easy to run in a container. ROCmFPX is now in the Lemonade Server from @AIatAMD

Welcome to the party!!
Halogen Flash is now a backend in @lemonade_server from @AIatAMD, just merged for Strix Halo: • Runs Qwen3.8-Flash-Next with v2 quants • Pinned to the Halogen 0.18.22 release • Tested in CI on the full model Thanks, Lemonade! github.com/lemonade-sdk/lemo…

🔥 GMKtec EVO-X5 Pro Is Here — First 30 Units from $6,399! Meet your next desktop AI workstation. ⚡ AMD Ryzen™ AI Max+ PRO 495 🎮 Radeon™ 8065S graphics 🧠 192GB LPDDR5X 8533MT/s memory 🤖 Run AI models with up to 320B parameters locally 🔗 Dual 10GbE LAN + dual 80Gbps USB4 ports 🔥 SUPER EARLY BIRD — first 30 units only: • 192GB RAM + 2TB SSD: $6,399 • 192GB RAM + 4TB SSD: $6,699 🎁 ZAI + 1.5× bonus points 📅 8-day launch offer: September 28–October 5 After the first 30 units, early bird prices apply: • 192GB RAM + 2TB SSD: $6,599 (regular $6,799) • 192GB RAM + 4TB SSD: $6,899 (regular $7,099) 🎁 $50 gift card + 1.5× bonus points 🛒 Shop EVO-X5 Pro: gmktec.com/products/gmktec-e… #GMKtec #EVOX5Pro #DesktopAI #LocalAI #AIWorkstation #AMD