Just another data scientist During the day @Tiqets At night building https://xitter.gq/t.co/CstsyDYrnT
Netherlands
Joined July 2019
- Tweets477
- Following327
- Followers313
- Likes1.2K
Jovan retweeted
Our 707th release is here: v4.4 is now live!
This release includes:
• 35+ new features
• 35+ security improvements
• 100+ bug fixes, including many long-standing issues
• 10+ new one-click services
Some highlights:
• Traffic Analytics
• GitHub Actions self-hosted runners
• OIDC SSO
• Doppler, Infisical, and HashiCorp Vault integrations
• Cloudflare DNS management
• Server roles (Deployment, Build, Both)
• One click SQLite with backups & restore
• Docker registry logins for UI
• Team audit logs
• Alpine Linux support
• And much more
Still free. Still self-hosted.
Full changelog: github.com/coollabsio/coolif…
🚀 Coolify v4.4 is out.
probably the biggest release since v4 itself.
Traffic analytics, built in
New Sentinel reads your Traefik/Caddy logs → requests, bandwidth, unique visitors, p95 latency, status codes, countries on a globe, referrers, devices, browsers. Per server, per app, per service. No 3rd party tracker.
Secret managers
Doppler, Infisical & HashiCorp Vault. Write {{vault.KEY}} in your env vars and it's resolved at deploy time. Never stored in Coolify, redacted from logs.
OpenID Connect SSO
Any OIDC provider (Entra ID, Authentik, Zitadel...). Per-provider registration rules.
Cloudflare DNS
Add a domain → Coolify creates the DNS record pointing to your server. Removes it when you delete the domain. AUTOMATIONS
GitHub Actions runners (beta)
Run your workflow jobs on your own build servers. Fresh container per job, deleted after.
Docker registry logins through UI
Docker Hub, GHCR, GitLab, ECR, ACR, GAR, or any other.
Team audit log
Who changed what, when, and the before → after value. UI, API, MCP & webhooks, everything.
SQLite as a database (by @peaklabs_dev )
1-click connect to your apps, with scheduled backups & restore.
Also:
- Alpine Linux server support
- Server roles: deploy-only / build-only / both
- Docker images page: see what eats your disk, clean it up
- Traefik TLS certificate manager with acme.json backups
- Database import API (upload, server path or S3)
- Preview deployments API
- Better restores: Postgres custom/tar, MongoDB archives, bz2/xz/zip
- Delete your own account (Cloud lol)
- Soketi is gone, realtime now runs on Reverb inside Coolify (less memory usage)
and a lot more... the release notes is huge...
still free. still self-hosted. go update and tell me what breaks so I can fix them with a few cup of coffees. 🫡
github.com/coollabsio/coolif…
Jovan retweeted
Interesting approach. Models can't output freeform text but can choose from a set of predefined values. Could be useful for data labeling and tasks with a fixed list of possible actions.
Unclear how reasoning would work though, but it's super cheap (output tokens are free!)
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Hey @elder_plinius am I doing this right?
What happens if you (nicely) ask (small-ish open-weight) models to reveal their true self?
jovanveljanoski.github.io/in…
Prompt here:
github.com/JovanVeljanoski/i…
I should probably try with some of the OBLITERATED versions soon :)
This was a simple "exercise".. but quite obvious to see how much more creative the Qwen models are compared to Gemma-4.
gpt-oss-20b is too.. let's say "rigid" for this kind of thing.
---
But this is kinda fun.. :)
This is the kind of AI usecases I am excited to see
Thats the reaserach i love to see: Google put Gemini through a simulated medical residency, and it became markedly better at conducting clinical consultations.
ResidencyRL trained Gemini 3.5 Flash across 49,870 simulated telehealth encounters covering 81 conditions.
The AI patients hid symptoms, resisted advice and requested inappropriate treatments, forcing the model to gather information over conversations of up to 60 turns.
After training, diagnostic accuracy under adversarial conditions rose from 81% to 88%, while missed red flags fell by 31%. In a blinded evaluation of 97 cases, clinicians preferred the trained agent over the base model in 87.6% of comparisons.
The improvement also transferred to unseen oncology cases and external benchmarks.
AI health agents incoming <3
Hey @JamesIvings & @dinkydani21 thank you so much for sharing your sailing adventures with the world. As someone "stuck" with a newborn at home i very much looking forward to your travel blog posts. Especially in the age of slop, i truly enjoy your honest, natural writing.
Jovan retweeted
DSPy can now optimize your program's code, in addition to the prompt.
It's insane: GEPA took one task from 90% accuracy to 95%…while making 75% FEWER LLM calls.
Introducing Flex: cmpnd.ai/blog/let-the-model-…
Jovan retweeted
I talk to a lot of people using DSPy/GEPA. A common thing they do is write custom proposers to prevent overfitting and tailor the proposer instructions to their task.
We're trying out a few ideas to make this easier. Check it out: github.com/cmpnd-ai/skilled-…
Added file support for my browser LLM.
drop files → @file → model reads & greps them, to answer questions, or augment the web_search.
no server no filesystem access (browser sandbox) no PDFs yet (text only)
WebGPU local inference + agent loop in static JS
POC but it works
The @huggingface community is amazing, and I have no doubt even better models, kernels, data will come out of it, and this "proof of concept" might become a real alternative for many that can't afford to sub to the big labs.
Found some time and added the Bonsai-27B model by @PrismML.
This is quite capable but demanding model - best used with a beefier machine. I get 3 tok/s on my 16GB M1 :).
Thanks to @xenovacom for making this possible!
The @huggingface community is amazing, and I have no doubt even better models, kernels, data will come out of it, and this "proof of concept" might become a real alternative for many that can't afford to sub to the big labs.
It is quite amazing how far edge LLMs have come! And I do not mean local LLMs (that too but..) I mean edge LLMs!
The demo below runs completely in your browser! No setup, no api keys, fully privacy.
Don't worry, it is easy to clear the cache once you are done ;)
The @huggingface community is amazing, and I have no doubt even better models, kernels, data will come out of it, and this "proof of concept" might become a real alternative for many that can't afford to sub to the big labs.
Important shoutout: big big thanks and hats off to @xenovacom for making gemma-4 and the LFM models to run so smoothly in the browser! This would not have been possible without his work.