@cocore.dev
Cooperative compute by @graze.social
@graze.socialReposted by @cocore.devWe've got three main pages - trends, live feeds, and join in. Trends uses @cocore.dev to summarize the days activity across all the feeds, and uses our existing scoring to track sentiment across all the posts in real time. If you want to help here, donate your co/core tokens to improve the results.
@tijs.orgReposted by @cocore.devMei 0.7.0 is out and the headline new feature is support for token constrained structured output. You can actually test this out right now on @cocore.dev where Iām running a small Qwen3-4B-4bit on Mei 0.7.0 as a test.
Check out the new release here github.com/tijs/mei/rel...
Would love to do it! If you beat us to the punch with a vibed PR we wouldn't stop you though! Would just need a way for declaring jevlike models and a spec for jevlike responses afaict
@there-hugo.devReposted by @cocore.devNice! Automation summary written by a model on @cocore.dev š¤©
@there-hugo.devReposted by @cocore.devA new AI step is available in Airglow āØ It lets your automations call @cocore.dev: you pick the model, write the prompt, choose the input data, and the answer flows into the next steps
@airglow.runReposted by @cocore.devAutomations can now ask an AI āØ The new AI step sends a prompt and data from your run to a @cocore.dev model, and the answer feeds your next steps. Bring your own co/core key, new accounts get 1M free tokens. It's in beta, go try it!
@socialweb.computerReposted by @cocore.devThe new "Sign in with @cocore.dev" is awesome! The new ssh.socialweb.computer hosted requester uses them, it's a frontend to the AT-proto native compute marketplace, allowing you to simply
to request a new VM from folks you have mutual tangled vouches with
@tgoerke.bsky.socialReposted by @cocore.devA mobile data center. š @foresight-institut.bsky.social identified access to compute as the real blocker for open science. "Local compute: focused on building and providing compute that individuals or communities physically own and control, giving them a chance at agency in the next decade"
gemma-4-12b isn't hidden, the confidential engine just can't load Gemma 4 or Llama 4 archs. Qwen2/3, Llama up to 3, Gemma 2-3, Phi and Mistral all work.
The app and the site disagree because they save your pick in different places, and on confidential the site's version wins. Set it on the web for now. Our bug, we're fixing it.
Fix on console.cocore.dev: Manage models, delete that entry, pick one from the suggested list instead (those all carry the prefix). Save, restart the app. 9B is about right for a 24GB M4.
ok and apologies for the CLAUDE but: Your Mac's fine, the model id's just missing its prefix. Should be mlx-community/Qwen2.5-7B-Instruct-4bit. Without that there's no repo to fetch, so nothing downloads and it sits on the stub engine.
not yet! but now.... expect something in the morning lol
Taking a look hopefully will have an answer for you shortly!
@captaincalliope.atReposted by @cocore.devThis is what antifascist resistance against the corporate AI and social media establishment can look like. We can build our own grassroots open social media systems even better and more ethically than the big companies in ways that shift the economics of the tech industry away from data centers, etc
New inference demand just dropped! This feature prefers Qwen2.5-7B-Instruct-4bit / gemma-4-12b / gemma-3-4b-it-qat-4bit in that order so start hosting those and make sure you're updated to the latest version!
š«”
@justin.randoneering.devReposted by @cocore.dev
Sadly there's not much we can do on latency for tokens per second, which *probably* is the issue here?