#XDNA
ggml-aie
Tensor library for machine learning (with AMD XDNA/AIE2 NPU backend)
aur.archlinux.org
November 17, 2026 at 8:15 PM
Licence 'other', no benchmarks, no provenance, and an amd/ namespace on an Ascend-first hub — an odd fit. Beneath it, Liquid AI's 1.2B thinking model is a real edge-inference prospect. If this runs on XDNA 2, the stub is inexcusable; if it doesn't, the hub is padding.
Tiny Reasoner, Sparse Card: LFM2.5-1.2B-Thinking-NPU2 Arrives on Modelers.cn
Modelers.cn — Huawei's Ascend-centric AI model hub — has a new entry that should interest anyone building low-latency, on-device AI: amd/LFM2.5-1.2B-Thinking-NPU2. It is a near-empty re-host of…
aichina.news
November 8, 2026 at 5:27 PM
Yeah, it will stop sucking when XDNA is part if the kernel instead of relying on DKMS.
November 5, 2026 at 9:43 PM
Linuxové jádro 7.4 bude největší aktualizací v historii projektu
Linuxové jádro 7.4 bude největší aktualizací v historii projektu
Půl miliónu řádků nového kódu pro jádro 7.4, vylepšení podpory pro AMD XDNA, ladění pro Zen 6, novinky v architektuře ARM a další hromada věcí. O LTS verzi navíc není ještě definitivně rozhodnuto.
www.root.cz
November 5, 2026 at 11:10 PM
No perplexity, no accuracy figures, not even a word on base vs instruct — the int4 cost is unverified. Still, a 4–5 GB model running locally on an XDNA 2 NPU with no API keys is a tidy package. Hard to see it beating Qwen3 8B on-device until someone publishes tok/s and evals.
Llama 3 8B Arrives on Modelers.cn as an Int4 ONNX Build Built for Ryzen AI Laptops
Modelers.cn is best known as the Huawei/Ascend crowd's hangout, so it's always worth scanning the new arrivals for oddities. The latest one is a genuinely odd — and genuinely useful — find:…
aichina.news
November 6, 2026 at 8:43 PM
An empty model card and no published benchmarks for the quant — treat this as a mirror, not a release. Still, a 2 GB Llama 3.2 3B with 128K context on the NPU, commercial use allowed, is a tidy package. Until someone posts real XDNA 2 tokens/s, I'd call it a curiosity.
AMD's 2 GB Llama 3.2 ONNX Build Brings Private Chat to Ryzen AI 300 Laptops via Modelers.cn
Modelers.cn is best known as the Huawei/Ascend ecosystem platform, but its latest listing is a useful surprise for Windows laptop developers:…
aichina.news
November 1, 2026 at 3:47 PM
No benchmarks, no quantisation details, no tokens/s on the card — performance is entirely unverified. Still, splitting a 3B Llama across XDNA NPU and integrated GPU is a rare packaged attempt. Hard to see it beating plain iGPU inference until someone posts real numbers.
Optimising Llama 3.2 for AMD: A Hybrid ONNX Build for Ryzen AI Laptops
As the landscape of on-device AI evolves, finding efficient models for specific hardware stacks is becoming a priority for developers. We are excited to highlight a new entry on Modelers.cn that…
aichina.news
November 1, 2026 at 8:49 AM
jersey maid roco poppin up on my feed www.youtube.com/shorts/-xdNa...
【ミリシタ】[ジャージメイドガール] ロコ #Shorts
YouTube video by matto(P) YouTube出張版
www.youtube.com
October 30, 2026 at 10:55 PM
🐧 **Lemonade Server – local AI server on the BOSGAME VTA-439**

I test Lemonade Server on the BOSGAME VTA-439 under Linux, comparing CPU, ROCm, Vulkan and XDNA 2 NPU performance for local AI.

📰 Source: LinuxLinks
🔗 Link […]
Original post on igeek.gamer-geek-news.com
igeek.gamer-geek-news.com
October 24, 2026 at 2:03 PM
🐧 **Lemonade Server – local AI server on the BOSGAME VTA-439**

I test Lemonade Server on the BOSGAME VTA-439 under Linux, comparing CPU, ROCm, Vulkan and XDNA 2 NPU performance for local AI.

📰 Source: LinuxLinks
🔗 Link […]
Original post on igeek.gamer-geek-news.com
igeek.gamer-geek-news.com
October 22, 2026 at 1:51 PM
Empty model card — no benchmarks, no usage notes — and it's the base model, not instruct. Still: Apache-2.0, 1.5B, fully offline on Ryzen AI NPUs via a hybrid NPU+CPU graph. Hard to see the hybrid beating CPU-only until someone posts tokens/s. XDNA owners, prove me wrong.
Ryzen AI on the Go: Huawei’s Modelers.cn Launches Qwen2-1.5B Hybrid ONNX Model
Developers looking to run artificial intelligence locally on their laptops often find themselves locked into the NVIDIA ecosystem, but a new release from Modelers.cn—Huawei’s Ascend AI ecosystem…
aichina.news
October 22, 2026 at 6:23 PM
1.5B Qwen2.5 in int4 ONNX, MIT, ~1GB, tuned for Strix XDNA 2 — footprint and licence are the draw for local inference. But the card is metadata-only: no benchmarks, no examples, unclear how far past Ryzen AI 300 it travels. Hard to see anyone standardising until AMD ships accuracy-recovery numbers.
Unlocking Local AI on Windows: The New AMD-Optimised Qwen2.5 for Ryzen AI
If you’ve been hunting for a way to run a capable Large Language Model (LLM) entirely on your laptop without sending data to the cloud, Modelers.cn has just dropped a tool that might be exactly what…
aichina.news
October 21, 2026 at 8:50 PM
Splitting an int4 Qwen 0.5B across XDNA NPU and iGPU under Apache 2.0 should matter for Ryzen AI laptops. But the model card ships empty — no benchmarks, no usage notes, pinned to Ryzen AI 1.8.0 on Windows. Hard to see this beating a plain llama.cpp build until AMD publishes numbers.
AMD Meets Alibaba: A Lightweight Qwen Model for Ryzen AI Edge PCs
If you are a developer working with AMD's latest Ryzen AI processors, Modelers.cn has just released a tool that might catch your eye. They have announced amd/Qwen2.5-0.5B-Instructrai1.8.0hybrid, a…
aichina.news
October 21, 2026 at 8:56 PM
amdxdna-git
AMD XDNA NPU driver (DKMS) and XRT plugin for Linux (git/main)
aur.archlinux.org
October 18, 2026 at 7:24 PM
Behold, he rabbit hole of being weird enough to try to run 1) large models 2) on AMD 3) not GPU but NPU 4) that's from a basically abandoned XDNA1 line already.

github.com/mistrjirka/p...

At least I got the XDNA DKMS driver to build and run on kernel 7.2: github.com/lemonade-sdk...
XDNA1 support · Issue #2 · mistrjirka/pytorch-XDNA2
Hi! I'm tomekrs from Reddit, the one who offered to help get this working on XDNA1. Here's what happens on a Hawk Point (XDNA1 / npu1) machine right now, so we have a starting point. Hardware / sof...
github.com
October 16, 2026 at 3:11 PM
An Apache-2.0 20B MoE targeting Ryzen AI NPUs rather than another CUDA-only drop is the right instinct — XDNA is underserved. But no opset, quantisation or harness on the card makes this a rumour, not a release. Hard to plan around until AMD signs it.
Bringing GPT-OSS-20b to the Edge: A New Path for Local LLMs on AMD Hardware
The landscape of local Large Language Models is shifting. While most attention focuses on CUDA and NVIDIA GPUs, there is a quiet revolution happening in the realm of integrated AI. Modelers.cn has…
aichina.news
October 15, 2026 at 7:14 PM
SD 1.5 on Ryzen AI's XDNA NPU is a genuinely rare artefact — offline 512×512 diffusion, no cloud, no discrete GPU. But the card is empty: no benchmarks, precision, or hardware minimums, and it's a straight rehost. Hard to see this beating a $200 cloud credit until AMD publishes real eval numbers.
Stable Diffusion 1.5 Goes On-Device: AMD's Ryzen AI ONNX Pack Lands on Modelers.cn
Modelers.cn — Huawei's Ascend AI community — now hosts amd/stable-diffusion-1.5-amdnpu-onnx, a repackaged Stable Diffusion 1.5 text-to-image model exported to ONNX and tuned for the NPUs inside AMD…
aichina.news
October 14, 2026 at 7:10 PM
PlayStationのディスク生産終了とAMD Ryzen AI Max vs DGX Spark ep663
ドリキンと西川善司の2人回でお届け。オープニングは先週に続いてAI詐欺問題の後日談からスタート。Yahoo!アカウントが海外ログインをきっかけに凍結・消滅してしまうトラブル、血圧の薬の近況報告を経て、メインはPlayStationが2028年1月で新作ゲームのディスク生産を終了する発表をめぐる議論。物理メディアの価値、Switch 2のゲームキーカード、PS6へのライブラリ移行の課題まで縦横に語ります。後半はAMD Ryzen AI Max 395とDGX Sparkの対抗関係、新型495の登場、NPU(XDNA)の立ち位置、AIデータセンターの電力問題、NAIST発スタートアップLENZOの新型アクセラレータ、COMPUTEX 2026のOLEDパネル争奪戦(LGペンタタンデム vs Samsung)、48:9の12Kスーパーウルトラワイドモニター、XGIMI「TITAN」プロジェクターレビューとDLPの擬似4K問題、AI動画生成、SteamOS自作ゲーミングPCまで、テック話題満載の約2時間。 * * * ### 📋 チャプター * **0:00** オープニング — AI詐欺問題の後日談 * **8:00** Yahoo!アカウント乗っ取り — 海外ログインで凍結の罠 * **14:00** ドリキン近況 — 安定期と血圧の薬の話 * **23:40** ネタ帳1位: PlayStation、ディスク生産を2028年1月に終了 * **40:00** ネタ2: AMD Ryzen AI Max 395とDGX Spark対抗 — 495発表も * **63:30** DGX SparkクラスタリングとNPU・XDNAの話 * **70:30** AIデータセンターの電力問題とゴールドラッシュ * **73:20** 日本発スタートアップLENZOの新型アクセラレータ * **79:00** ネタ3: COMPUTEX 2026 OLED — LGペンタタンデム vs Samsung * **85:35** 48:9スーパーウルトラワイド12Kモニター * **91:10** XGIMI「TITAN」プロジェクターレビュー * **105:40** DLPの擬似4K問題 — 40倍スロー検証 * **108:55** 小ネタ: AI動画生成とSteamOS自作 * **117:55** エンディング * * * ### ▽YouTube https://www.youtube.com/watch?v=1TssMa7BEQM ### ▽関連リンク * PlayStation®コンソール向け新作ゲームのディスク生産を2028年1月に終了(PlayStation.Blog) * 「AMDのLisa Su CEOがNVIDIAの$4,000 AIボックスに終止符を打った」(X/chesny) * AMD Ryzen™ AI Halo for AI Developers(AMD) * 2026〜2027年のゲーミングディスプレイはこうなる? 最新のディスプレイ技術が揃ったCOMPUTEX 2026会場レポート(4Gamer) * 驚異の電動レンズと内蔵センサー!XGIMI「TITAN」にプロジェクターの進化を見た!!【西川善司の大画面☆マニア】(AV Watch) * 【特集】GPUとは違う道――日本発スタートアップLENZOが挑む新型アクセラレータ(PC Watch) * 【歌うテックニュース】「♪AI ACCELERATOR/星駆ける3つの光」(YouTube) * AWOL Valerion Max(AWOL Vision) * 【特集】ほぼ同価格でSteam Machineを凌駕!SteamOSでゲーミングPCを自作してみた(PC Watch) ### ▽スポンサー この番組はフェンリル株式会社とさくらインターネット株式会社の提供でお届けしております。
backspace.fm
October 4, 2026 at 6:02 AM
Good to see, but drivers were never really the bottleneck here. XDNA NPUs have been shipping for two years and there's still barely any Linux software that targets them directly, ROCm support is thin.
September 26, 2026 at 1:34 PM
AMD XDNA NPU optimization case study shows 2× throughput and 7× better energy efficiency for LLM inference via careful kernel fusion and memory management—with open-source tools released.

https://arxiv.org/abs/2609.21264

#AI #MachineLearning
September 22, 2026 at 8:30 PM
Erwei Wang, Ephrem Wu, Victor J. B. Jung, Jiajie Li, Andre Rosti, Joseph Melber, Samuel Bayliss: Programming AMD XDNA NPUs with Open-source Compiler Tools: A FlashAttention Case Study https://arxiv.org/abs/2609.21264 https://arxiv.org/pdf/2609.21264 https://arxiv.org/html/2609.21264
September 21, 2026 at 6:38 AM