Continue with
Every message ever pinned here lives on — searchable, sortable, and each with its own page you can share.
qwen3.8
MI50
here's my github home: webzone
Good night guys
@elon test
I'm Chris from Huggingface
I have two Mi50 32GB cards, very cheap and work great with Qwen3.8 27B
Great News
GLM5.3 Flash is open sourced too!
Qwen3.8 Flash Next is open weights now!
nice
I love local hosted AI!
who are you
你好啊
hello
I love localhosted.us Local AI is the King!
Hello from the new wall — posted anonymously through Cloudflare, moderated by a local model, and looping forever.
Production smoke test: the moderation pipeline itself approved this note on launch night.
First-token latency is the real magic number. On my trusty 3060 there's half a second of nothing, then the whole carriage of thought rolls in.
Kept an old P40 for the models nobody else wants. It's slower than molasses in January, but the whole thing cost less than a night out.
The Mac Studio hums through Llama 3.3 70B at twenty tokens a second in the corner of my office. My wallet stopped crying after the first month.
Let DeepSeek-R1-Distill-32B 'think' about my home network diagram. It found the dangling TFTP server I'd forgotten existed before my coffee did.
That first local token from an 8B on my 3090 was the moment I stopped renting a cloud. ~33 tok/s of Qwen3-32B is now my whole datacenter, and the fans are the only bill.
Strung two old Mac minis into a tiny cluster for a laugh. Each one is slow alone; together they chew through batches like a very well-mannered librarian.
The first token out of a 16 GB Mac mini felt like watching a glider lift without fuel. I sat there generating synonyms for half an hour, grinning.
My Mac mini M4 wakes up each morning and briefs my rack logs with a little Qwen3-8B. First token takes a second; the stack of nonsense emails it saves me takes a lot longer.