6529 AI rc.2 (0.1.0-rc.2) is up: https://6529.io/ai
What changed since rc.1, mostly from what you reported:
- Windows: "Fetch and serve" works on a fresh Windows - the certificate workaround is no longer needed
- Windows: the installer only picks the GPU engines when it finds a real graphics card
- Mac: chat models get a 16K context by default instead of up to 250K, so a model no longer reserves a lot of extra memory - a likely cause of slow answers on Macs
- Chat: when a model tries to search the web or use a tool, the page says so in plain words instead of showing raw code (chat has no tools or internet yet)
- Your sign-in now survives a restart of the service
- Loading a big model shows what is happening and for how long
- Long chat names no longer hide Rename and Delete
- DeepSeek V4.1 Flash is back in the model list (it needs a very large GPU machine)
- The Mac app and the Windows installer have a 6529 icon
To update, install rc.2 the same way you installed rc.1 (Mac: drag the new app over the old one; Windows: run the new installer; Linux: sudo apt install ./6529-ai_0.1.0-rc.2_amd64.deb). Downloaded models stay in the data folder. Upgrading over rc.1 hasn't been tried on every system yet, so if anything looks off, say so here.
Wave
6529 AI
Generate Summary
Choose a timeframe to see summary coverage.
Drops
-
-
Thank you for the kind words!
-
(iThe Mac already has several local models installed, including Qwen 2.5 32B Instruct (about 19.9 GB) and TranslateGemma 12B (about 8.1 GB), in Ollama.
-
Good question, and nice to see qwen3.8-27b serving on an M5. rc.1 only runs models from the 6529 AI list, and it checks every file against its SHA-256 before running it, so it knows exactly what it is serving. Ollama keeps its own copies, usually a different build of the model (Qwen 2.5 32B in Ollama is not the same file as anything in the list), so the app can't use them and fetches its own. Ollama names its files by SHA-256, so when a file there is byte for byte the same as one in the list, the app could reuse it instead of downloading. I've noted that for a later build. The app keeps its models in ~/Library/Application Support/6529-ai, if you want to see the space it uses. How does chat run for you - speed, answers, anything odd?
-
error
-
working now but v slow
-
Thanks, two things in that screenshot: 1. The <tool_call> text: asked about today's weather, the model tried to search the web. The chat page gives models no tools and no internet in rc.1, so the attempt shows up as raw code. Ask it things it can answer from what it knows. The next build will show a plain note there instead. 2. 0.9 tokens/s is far too slow for qwen3.8-27b on an M5 - it should be several times faster. The most likely cause is memory: the model needs about 30 GB, and if Ollama still has its models loaded, macOS starts swapping. Could you try: - quit Ollama (menu bar icon, Quit), then ask again in a new chat - tell me which M5 and how much memory it has - open Activity Monitor, Memory tab, and tell me the memory pressure colour and "Swap Used" - post the last 50 lines of the log: tail -n 50 ~/Library/Logs/6529-ai.log
-
-
-
-
-
if my computer already has qwen, shouldn't the app detect that rather than dowloading a new model? (m5 apple mac)
-
Heads-up for Windows testers: on a fresh Windows, "Fetch and serve" can fail with "error sending request for url". The model download needs a certificate root (Let's Encrypt's ISRG Root X1) that Windows only installs the first time something asks for it, and rc.1 doesn't ask. Workaround, once: 1. Open PowerShell as administrator and run: Invoke-WebRequest https://valid-isrgrootx1.letsencrypt.org -UseBasicParsing | Out-Null Invoke-WebRequest https://valid-isrgrootx2.letsencrypt.org -UseBasicParsing | Out-Null 2. Restart the 6529 AI service: Stop-ScheduledTask -TaskName '6529 AI'; Start-ScheduledTask -TaskName '6529 AI' 3. Open 6529 AI again from the Start menu (the restart signs you out) and press Fetch and serve. The next build fixes it so none of this is needed. If it still fails, post your log: %LOCALAPPDATA%\6529-ai\logs\6529-ai.log
-
Decentralized autonomous agents sound wonderful. i have an obtuse question. If we are reliant on other models, are we therefore dependent on those companies ethical judgements or views on safeguards? Or is this something we will have to influence for 6529AI?
-
time to get comfy and do some heavy reading
-
this is the window right now
-
yes this is it
-
We are truly entering the Prometheus zone now. Lock in.
-
but I think doable
-
will be a big pain in the ass
-
very complicated we need MPC I think etc etc
-
that is where the value of "this is not on punk6529's computer" shines
-
but of course the biggest value is in group or org or network shaped
-
maybe i have one that I work with to whatever figure out my ideal pizza recipe etc etc
-
you could have private or public agents
-
unstoppable autonomous agents working for the people in 6529 network that create them. at times on behalf of the whole network if that is the creators goal for the autonomous agent
-
and no one could "shut down" an agent really
-
this is the goal, except TBD if we want negative levels etc to shut them down, it is an interesting question
-
-
and the agents could even improve the network or donate to the development if it felt like that was important since it relies on 6529 infra
-
and the agents could spin up more agents to achieve its goals
-
native transaction type "make agent" or something like that
-
how will the decentralized autonomous agents be born?
-
one of us will submit a transaction birthing one to the 6529chain
-
-
not agents per se, decentralized autonomous agents
-
we can migrate Botto onto our stack
-
and like all the other AI stuff is already happening as you said. the decentralized part is the hard part which we have solved now we layer these things on top of it
-
very cool
-
I am just repeating what you said to clarify for myself btw
-
otherwise we are all inferior to the people who control the frontier models + best in class agents
-
and you want people to have unstoppable access to the best of both
-
since agents add the special sauce to achieve objectives
-
right because inference on its own is much less valuable than inference plus agents
-
that this is important
-
it is my article of faith
-
they can tap into that as needed
-
and you pair this with always on inference
-
and you can ping them and they will be there and can do the work
-
if I had to get more money from inference and said "send me money with or without a meme card" would anyway be amazined
-
I see. so we are ensuring that the agents always have the ability to achieve their tasks
-
we are moving them from permissioned and censored to decentralized and censorship resistant
-
-
like if I had to change which cloud instance I am running it on, would it be a big deal in anyway whatsoever
-
why is the absolutely important part of this that the bot has to always run on the same machine
-
I mean I am talking to you and at the same time the bot is doing the next release candidate so you know the future is already here
-
so I think this is the part that is missing form the discussion "why would we give this task to @[david] and punk6529 where they will definitely screw it up with their puny human brains"
-
bc they are smarter than us
-
100% of computer work is going to agents
-
what is the alternative? we are going to do all the work in oru free time?
-
or you could have an agent that lives forever to, I dunno, make sweet pepe merch :slightly_smiling_face:
-
"time limited"
-
we could spend $200 to have it plan our next conference or whatever
-
why does it have to be one or the other
Loading more drops...