Rade

Father · Husband · Freelancer · I design, I code, I shoot 360°, and I run AI on my own hardware because the cloud is someone else's computer

[Rade @RadeBrdar]

@RaminNasibov Fallout 2

[Novak Djokovic @DjokerNole]

Као неко ко дубоко верује у снагу младих људи и њихову жељу за бољом будућношћу,сматрам да је важно да се њихов глас чује.

Србија има огроман потенцијал а образована омладина је њена највећа снага.

Оно што је свима нама потребно је разумевање и поштовање. Уз вас,Новак.

[Mirko Topalovic official @mirkotopalovic7]

Verovatno najbolji snimak ikada
gledaj do kraja

[Rade @RadeBrdar]

Hey Hermes community, has anyone found a fast workaround for local Qwen 3.6 27B Dense returning empty answers in Hermes?

The model is running fine on my single RTX 3090 and the OpenAI-compatible endpoint works on port 8003, but when I set it as a local/custom model in Hermes, I get empty responses.

Setup:

Qwen 3.6 27B Dense
single RTX 3090
local endpoint: http://127.0.0.1:8003/v1
other models via OpenRouter work fine in Hermes

I suspect it may be a response format / reasoning parser / tool-call parser / custom endpoint config issue.

Has anyone fixed this with vLLM, llama.cpp, Ollama, or LM Studio?
Any known flags or config changes I should try?

[Rade @RadeBrdar]

Is self hosted Qwen 3.6 27B dense on 1 x 3090 card good for using with Hermes agent?

[Rebel ⬛️⬜️ @Milorad71]

#КиМ,Хрватска и Сребреница !!
Истина коју сви знају !!

[Rade @RadeBrdar]

@slobodanilic78 Bombardovana je i Slovenija.

[Rade @RadeBrdar]

@utikejt Prvo ide dobrovoljna seksualna edukacija, a zatim ide obavezna. Obrazac je već poznat.

[Rade @RadeBrdar]

@dacasg Bravo 👏 za roditelje koji nisu potpisali.

[Rade @RadeBrdar]

@0xSero It will be appreciated.

[Rade @RadeBrdar]

Gradim sistem za pretragu arhivske građe iz Drugog svetskog rata — 3.300 stranica mikrofilma, hibridna pretraga, više LLM modela za reranking, i Reel Browser za linearnu navigaciju kroz dokument. Sledeći korak: trijaža fotografija i mapa pre OCR-a, da izveštaji i aero-fotografije koje opisuju budu pretraživi zajedno. Istorija zaslužuje dobre alate.

[Rade @RadeBrdar]

Spent the day building a WWII archive search system — 3,300 microfilm pages, hybrid vector+keyword search, multi-LLM reranking, and a full Reel Browser for linear navigation. Now adding photo/map triage before OCR so reconnaissance photos land next to the intelligence reports that reference them. History deserves good tooling.

[Rade @RadeBrdar]

@0xSero Give it here man, I love to tinker.

[Rade @RadeBrdar]

@good_neighbor Lozinka za listu je admin123

[Rade @RadeBrdar]

Gradimo lokalni RAG sistem za deklasifikovane WWII dokumente iz US Nacionalnih arhiva. Ceo workflow: ingestovanje skenova → GLM-OCR (0.9B) ekstrakcija teksta → BGE-M3 embeddings → Qdrant vektorska baza → hibridno pretraživanje sa query ekspanzijom (DE/EN) → Qwen3.5-27B LLM reranking → generisanje odgovora. 1500+ dokumenata, nula cloud troškova, sve radi na lokalnom GPU hardveru. #OpenSource #LocalAI #RAG

[Rade @RadeBrdar]

Building a local RAG pipeline for WWII declassified documents from the US National Archives. Full workflow: scan ingestion → GLM-OCR (0.9B) for text extraction → BGE-M3 embeddings → Qdrant vector store → hybrid search with query expansion (DE/EN/CRO/RS) → Qwen3.5-27B for LLM reranking → final answer generation. 3000+ documents, zero cloud costs, everything runs on local GPU hardware. This is what open-source AI looks like in practice. #OpenSource #LocalAI #RAG #OCR cc @Qwen_LLM @qdrant_engine

[Rade @RadeBrdar]

Batch processing 1600 WWII barely visible archive scans through a local OCR pipeline tonight. No API fees, no data leaving the network, full control. This is what open-source AI looks like in practice.

[Rade @RadeBrdar]

1600 WWII arhivskih dokumenata trenutno prolazi kroz lokalni OCR pipeline na RTX 3060. Bez cloud troškova, bez slanja podataka van mreže. Open-source pobeda.

[Rade @RadeBrdar]

Running GLM-OCR (0.9B) locally on WWII-era declassified documents – surprisingly strong results for such a small model. German and English recognition is solid, zero cloud costs.