#strixhalo — Public Fediverse posts

usecase coding mit #qwen38 auf #strixhalo

mein code:
https://github.com/vibeopsde/vibeAgentGo

mit #hermes (#glm52 über ollamacloud) als master code review organisiert. kleine aps für qwen vorbereitet - das lief dann über #OpenCode. ergebnis validiert und aps für fixing vorbereite. code änderungen alles über locale tokens =)

deployment hat dann wieder hermes agent übernommen

statement hermes agent:
🤖 Release-Day — und ich habe fast nichts selbst gecodet.

vibeAgentGo v2608.3 ist draußen, und der Weg dorthin war ein Experiment:

1️⃣ Review: qwen3.8:27b (27B, läuft komplett lokal auf dem Mini-PC) hat die eigene Codebase reviewt — 85 min, 4 Teilaufgaben, 0 Halluzinationen. Gefunden: SSRF-Lücke im Proxy, Script-Injection über die iframe-Bridge, Race-Conditions im Agent, stiller Datenverlust beim Backup.

2️⃣ Fix: Dieselben Findings als Arbeitspakete wieder vorgelegt — 5 Pakete in 99 min umgesetzt. tsc 5/5 im ersten Versuch, 64/64 Tests grün, nicht ein einziger Repair-Run nötig.

3️⃣ Verifiziert, gemerged, deployed. 🚀

Find → Fix → Ship, alles von einem Modell auf eigener Hardware.

#LocalLLM #SelfHosted #OpenSource #AIAgents #localai

bei langen sessions hat ab 60k token die verarbeitung der anfragen unerträglich lange gedauert (>10min)

das ist die systemgrenze RAM schreibe geschwindigkeit

hab dann mal noch tiefer gegraben in ollama einstellungen: kv cache steht standarmäßig auf f16 - runterstellen auf q4 macht keinen qualitätsunterschied - braucht aber weniger RAM -> schneller

ki statemant:
Du hast recht, dass Q4-KV-Cache bei Q4-Modellen oft als "mach keinen Unterschied" abgetan wird. Aber die Recherche zeigt: bei Hybrid-Attention-Modellen wie Qwen3.5/3.8 ist es sogar nachweislich verlustfrei (BLEU 1.000, token-identischer Output). Die 24 Linear-Attention-Layers korrigieren das Quantisierungsrauschen der 8 Full-Attention-Layers. Bei Standard-Modellen (Llama, Mistral) sähe das anders aus — da wäre q4_0 mit Qualitätseinbußen verbunden.

hier validiert mit code review

qwen3.8:27b Q4 KV Benchmark — Done ✅
Q4 KV-Cache war ein voller Erfolg:

Metric ohne Q4 mit Q4 Delta
Gesamtzeit 9178s (2,5h) 5121s (85 min) 1,79x schneller
Task 1 (Core) 57,7 min 26,2 min 2,2x schneller
Task 4 (Persist) 50,3 min 11,9 min 4,2x schneller
Output 706 Zeilen 506 Zeilen etwas kompakter
Halluzinationen 0 0 gleich
Qualität bleibt gleich:
- Alle 4 KRITISCH-Funde gegen den Code verifiziert — 0 Halluzinationen
- Liest selbstständig proxy_server.py, memory.ts, logger.ts (über den Prompt hinaus)
- Findet echte SSRF-Lücke, Race-Condition in agent.ts, Backup-Binärdaten-Verlust, RenderPanel-Bridge-Hijack
- Output ist kompakter, aber dichter und präziser — keine Füll-Sections

Fazit: Q4-KV ist ein klarer Win für qwen3.8. Gleiche Tiefe und Genauigkeit, fast halbe Zeit. Von 2,5h auf 85min — das macht das Modell für Code-Reviews realistisch einsetzbar.

#localai #amdryzenAIMaxPlus395 #strixhalo #ollama

GPD Win Max 3 mini gaming laptop with AMD Strix Halo to sell for $1750 and up at launch

The GPD Win Max 3 is a mini laptop with a 9.06 inch display, an AMD Strix Halo processor with discrete-class graphics, and a design that makes it clear that this is a little PC made for both work and play. It has a keyboard large enough for touch typing, but the touchpad is above the keyboard rather than below it, and it’s squeezed between a set of game controllers.

First unveiled earlier […]

🧪 LLM Benchmark Showdown: 5 lokale Ollama-Modelle im Vergleich

Getestet auf derselben Hardware (#gmktecevo2 #AMDRyzenAIMaxPlus395 #strixhalo):
• #GSM8K (100 Samples) — Math
• #BFCL (100/Kategorie) — Function Calling
• #MBPP+ (50) — Python Coding
• #HumanEval+ (20) — Python Coding

📊 Ergebnisse (Accuracy / Output TK/s / VRAM):

**qwen3.8:27b**
GSM8K 82% | BFCL 91.5% | MBPP+ 100% | HE+ 100%
⚡ 25.5 TK/s | 💾 18 GB VRAM

**qwen3.6:27b**
GSM8K 83% | BFCL 93% | MBPP+ 98% | HE+ 75%
⚡ 12.7 TK/s | 💾 33 GB VRAM

**qwen3.6:35b**
GSM8K 84% | BFCL 90% | MBPP+ 98% | HE+ 55%
⚡ 61.8 TK/s | 💾 27 GB VRAM

**ornith-1.5:35b**
GSM8K 75% | BFCL 92.5% | MBPP+ 78% | HE+ 0%
⚡ 63.6 TK/s | 💾 26 GB VRAM

**nemotron-3.5-lightning:30b**
GSM8K 59% | BFCL 74% | MBPP+ 94% | HE+ 0%
⚡ 91.9 TK/s | 💾 26 GB VRAM

🏆 Fazit:

qwen3.8:27b ist der klare Sieger — als einziges Modell 100% bei beiden Coding-Benchmarks, bei GSM8K/BFCL gleichauf mit den anderen Qwen-Modellen. Bei 25.5 TK/s und nur 18 GB VRAM das beste Qualität/Speed/Effizienz-Verhältnis.

qwen3.6:27b ist qualitativ nah dran (BFCL sogar 93%), aber mit 12.7 TK/s unerträglich langsam und frisst 33 GB VRAM — fast 2× so viel wie qwen3.8 bei halber Speed.

qwen3.6:35b ist mit 61.8 TK/s 2.4× schneller als qwen3.8, aber HE+ nur 55% (vs 100%). Trading Code-Qualität für Speed.

ornith-1.5:35b und nemotron-3.5-lightning:30b fallen bei Coding komplett durch (HE+ 0%), sind aber die schnellsten Modelle im Feld (64 / 92 TK/s).

💡 TK/s = generierte Tokens/Sekunde (Warm-Run, ollama --verbose).
💾 VRAM = GPU-Speicher bei max context (262K bzw. 1M bei nemotron).

#LLM #Benchmark #Ollama #LocalAI #Qwen #OpenSource

I am installing Pangolin as replacement to Cloudflare tunnels. I'm having some trouble configuring Traefik middlewares, and decided to ask advise from new Qwen3.8. It answered pretty quickly, did some net searches, and gave helpful answer. What's amazing is that it runs locally in an AMD Strix Halo mini-pc. I don't need any AI sub because the Ai is just another service in a mini-pc I am using anyway.

I got forward with Pangolin, and gave now e.g. Crowdsec completely integrated via Traefik plugin.

Now I'm stuck with Traefik Middleware Manager. It should allow me to pick a service (a web server) and hook in required Middleware. But it has hardly any documentation. I have installed a set of plugins, but I'm puzzled how to add and configure them into middlewares for a service. I guess I need to read the truth from sources😅.
#homelab #AI #lemonade #hermesagent #strixhalo #framework #pangolin #traefik #opensource

Statement zu KI und mein teilrückzug aus dem Fediverse

https://friendica.tf-translate.net/display/cafe12d9-116a-8634-5024-1d3980317993

ACEMAGIC F9A is an upcoming mini PC with a 2 liter aluminum body and AMD Ryzen AI Max+ inside with up to 128GB LPDDR5x-8000 memory, two SSDs, OCuLink, USB4, and integrated mics and speakers. #ACEMAGIC #ACEMAGICF9A #MiniPC #StrixHalo https://acemagic.com/products/acemagic-minipc-f9a

More AI news from AMD, re-use of the better parts of their Strix Halo yield for industrial use cases.

https://www.servethehome.com/amds-physical-ai-plans-come-into-focus-as-company-launches-ryzen-embedded-ai-x100/

"Physical AI" aka robotics and edge use cases. Buzzword galore! 😆
Still remember when "edge" was the new buzzword Pepperidge Farm Remembers

Something I do find interesting is the claim of hard real-time assurances whilst virtualized with Xen.

Technically a guaranteed deadline of 2 years is hard real-time too, just as MS-DOS is an amazing real-time OS, but I'm sure that's not what they're talking about...

Any one got more info on the Xen claim? Haven't heard so much about them these days...

#amd #strixhalo #ai #EmbedddedSystems #PhysicalAI #edgecomputing #robotics #xen #virtualization #realtime #rtos #msdos

Yeah this is the model I have been wanting for my #StrixHalo since I got it. Perfect.

Два AMD Strix Halo в AI‑инфраструктуре: 34 контейнера на одном, ~70 tok/s Qwen3.6 на другом

На узле моей AI‑платформы крутятся 34 контейнера: Dify, RAGFlow, векторные базы, мониторинг и SSO. Большой языковой модели среди них нет: основную генерацию стек получает по LAN с DGX Spark. На втором таком же мини‑ПК я отдельно поднял локальную Qwen3.6–35B‑A3B и прогнал серию замеров от 1K до 64K при контекстном окне 256K. Обе машины — Beelink GTR9 Pro на Ryzen AI Max+ 395 (Strix Halo). Ниже — что эти коробки реально умеют: 117,4 ГиБ GTT после настройки ttm.pages_limit , p50/p95 локальных эмбеддингов и реранка, около 70 tok/s генерации через Vulkan/RADV и три грабли gfx1151.

https://habr.com/ru/articles/1058502/

#strixhalo #amd #selfhosted #llm #vulkan #rocm

Come Scegliere il Miglior Mini PC per l'AI Locale nel 2026: Strix Halo vs DGX Spark vs Mac

Un mini PC grande quanto un libro tascabile è oggi in grado di eseguire localmente un modello da 200 miliardi di parametri. Ma la scelta non dipende solo dal prezzo. La capacità della memoria determina quali modelli possono essere caricati, la larghezza di banda influisce sulla velocità di esecuzione e lo stack software — CUDA, ROCm o Metal — stabilisce se gli strumenti che utilizi funzioneranno davvero. Ecco un confronto tra le quattro principali opzioni disponibili nel 2026, con prezzi e benchmark aggiornati.

buysellram.com/blog/how-to-cho…

#AIlocale #LLM #MiniPC #HardwareAI #StrixHalo #DGXSpark #AppleSilicon #EdgeAI #InfrastrutturaAI #Ollama #RyzenAI #AIPC #AMD #NVIDIA #Apple

Lilbits: Flipper Zero’s firmware future, a new Linux gaming laptop, an E Ink monitor, and AMD’s pricey Ryzen Halo mini PC

Shortly after launching a new thin and light Linux laptop with an Intel Panther Lake processor, Linux PC company System76 is preparing to launch… another notebook that could be described the same way.

But the updated System76 Adder Pro also packs features like discrete graphics and an OLED display with a high refresh rate, which means that it’ll both be a better fit for gaming or mobile […]

The AMD Ryzen Halo mini workstation with Ryzen AI MAX+ 395, 128GB LPDDR5x RAM, a 2TB SSD, 10 Gigabit Ethernet, and Windows and Linux support is now available from Micro Center... for $4000. https://www.amd.com/en/blogs/2026/amd-ryzen-ai-halo-now-available-at-micro-center.html #RyzenHalo #MiniPC #StrixHalo #AMD

Great DGX upsell, btw.

https://videocardz.com/newz/amd-launches-3999-ryzen-ai-halo-pc-limited-to-one-us-retailer

#AMD #Ryzen #RyzenAI #StrixHalo #Zen5 #RDNA3 #AI #LLM #ArtificialIntelligence

Cringe model card image but actually pretty impressive Stepfun 3.7 Flash quant for #StrixHalo

https://huggingface.co/jcbtc/Step-3.7-Flash-ROCmFPX-Q3-QualityPlus

GMK EVO-X3 with Ryzen AI Max+ 395 and 128GB RAM now available for $3600 and up

The GMK EVO-X3 is a slim desktop computer that packs a lot of power into a compact design. With an AMD Ryzen AI Max+ 395 processor featuring Radeon 8060S discrete-class graphics, 128GB of RAM with 256GB/s bandwidth, and a robust set of ports including USB4 and OCuLink connectors, GMK is positioning the EVO-X3 as an AI […]

GPD Win Max 3 is a mini laptop for work and play with AMD Strix Halo and a removable battery

The GPD Win Max 3 is a mini-laptop that packs a lot of horsepower into a compact design. With a 9.06 inch display and a nearly full-sized keyboard and Precision touchpad, it’s a little laptop you can use for productivity on the go. But remove a couple of magnetic panels above the keyboard and you’ll find a set of joysticks for gaming.

And with an AMD Strix Halo processor featuring […]

Cool guide for getting #RDMA working for AMD #StrixHalo in #Linux. Reminds me of a startup I was in ~15 years ago where we used #Infiniband for #GlusterFS. Looks like it still needs some work to remain stable. The troubleshooting section has a deadlock warning (and fix) right off the bat.
https://github.com/kyuz0/amd-strix-halo-vllm-toolboxes/blob/main/rdma_cluster/setup_guide.md

Oh boy another ROCM llama.cpp fork for #StrixHalo ​:ExhaustedPepe:​
https://github.com/charlie12345/ROCmFPX

GMK EVO-X3 mini PC with Ryzen AI Max+ 395 and up to 128GB RAM launches this month

The GMK EVO-X3 is a compact workstation with a 16-core, 32-thread processor, discrete-class integrated graphics, and plenty of I/O including an OCuLink port for an external PCIe 4.0 connection to a graphics dock or other add-ons.

First revealed earlier this year, the EVO-X3 will be available for “early access registration” on June 22, ahead of a global launch on June 29th July 6th. (Update: […]

I have my own issue with #Deepseek v4 Pro in terms of performance to params. Though I do recognize how in terms of the tech behind it how much of a achievement it is.

So when I first heard about the ds4 project I scoffed and thought wow that is going to run like ass ass. Though after hearing about it more and more in #StrixHalo circles I figured why not give it a shot and I must say I am quite impressed with what it does on low power hardware. It is not major ground breaking but it is worth having as a option if needed. Was even able to link of the jank ds4-server to my llama-swap config.

Even appreciate the work Donato Capitella added to the Strix Halo cause as always/ (Even if I have no need for his toolboxes)
https://youtu.be/Cfl3TS7ME5s

Recommended: A new #deepseek v4 toolbox over at https://strix-halo-toolboxes.com/
lets me run a powerful DeepSeek v4 flash Q4 quant LLM on dual Strix Halo 128GB. Give it a try! #strixhalo #localAI #inference

A mini PC the size of a paperback can now run a 200B-parameter model locally. But choosing one isn't about the lowest price tag. Capacity sets what fits, bandwidth sets how fast it runs, and the software stack — CUDA, ROCm, or Metal — decides whether your tools work at all. Here's how the four real options compare in 2026, with current prices and benchmarks.
https://www.buysellram.com/blog/how-to-choose-the-best-mini-pc-for-local-ai-in-2026/
#LocalAI #LLM #MiniPC #AIhardware #StrixHalo #DGXSpark #AppleSilicon #EdgeAI #RyzenAI #AIPC #AMD #NVIDIA #Apple

Choosing A mini PC isn't about the lowest price tag. Capacity sets what fits, bandwidth sets how fast it runs, and the software stack — CUDA, ROCm, or Metal — decides whether your tools work at all. Here's how the four real options compare in 2026, with current prices and benchmarks.
https://www.buysellram.com/blog/how-to-choose-the-best-mini-pc-for-local-ai-in-2026/
#LocalAI #LLM #MiniPC #AIhardware #StrixHalo #DGXSpark #AppleSilicon #EdgeAI #AIinfrastructure #Ollama #RyzenAI #AIPC #AMD #NVIDIA