Umans AI

Building intelligent systems that help people deliver software that actually solves problems.

[Wassel @wasselovski]

This how DeepSeek V4 Flash feels @umans_ai ⚡️

[Wassel @wasselovski]

Kimi K3 is landing on Umans AI.
It’s the largest frontier model we’ve operated so far, served with zero data retention.
Real Open AI is winning. 🙂
And we’re only getting started.

[Wassel @wasselovski]

Tomorrow we’re opening Kimi K3 to the community for a new experiment.

Before that, two short write-ups from models we tested 👇

[Umans AI @umans_ai]

@Thewimo @Kimi_Moonshot since few hours ago
https://t.co/Kx0m4iMMmj

[Wassel @wasselovski]

Kimi K2.7-code is now available to all @umans_ai subscribers.
https://t.co/mTJzoK0fUn

[Wassel @wasselovski]

Just shipped: Kimi K2.6 and GLM 5.1 on Umans Code.
Open models keep catching up.
We pick the best for coding and run them on our own infra.
Still iterating. Tell us how it works for you!
https://t.co/pvfc8OrbOQ

[Wassel @wasselovski]

Wrote up how we actually ship with ai at @umans_ai https://t.co/wYDub0FzGg

[Wassel @wasselovski]

Started using @zeddotdev recently and honestly I like it a lot.

Clean, fast, low distraction.

Quick video on how to plug it into Umans Code for unlimited tokens

[Wassel @wasselovski]

One week after launch, the signal is pretty strong.
Umans Code is already seeing strong real-world usage.
Since Feb 6:
- users from 50 countries
- 81.5% daily active users over the last week
- top countries by users: France 🇫🇷, USA 🇺🇸
, India 🇮🇳 ❤️

Builders want to stay in flow and let their agent keep working.
That is exactly what we are building.

Thanks to everyone building with it and helping us improve it.

[Wassel @wasselovski]

@codewithsushi Builders should move to https://t.co/6loEFZVyQD
To keep building without getting stressed about Anthropic bill

[Wassel @wasselovski]

Can’t agree more!
This is why we kept Kimi K2.5 our default model at https://t.co/6loEFZW6Gb
Unlimited tokens 50$ p/m

[Wassel @wasselovski]

MiniMax M2.5 is now live on Umans Code (umans-minimax-m2.5).
Vision + web search, ready to use with Claude Code. No extra setup. No MCP.
Users are testing it right now. It may become our default model, for now Kimi K2.5 stays the default.

[Wassel @wasselovski]

@iammstachowiak @Kelly__cloud @antirez Here is how to use it with cursor.
DM if you need any help.
Good luck with launch next week 🚀

[Wassel @wasselovski]

We just shipped the Umans CLI 🚀
It lets you run Claude Code with an unlimited plan powered by the best OSS coding model of the moment.
Demo video 👇

[Wassel @wasselovski]

I kept saying “available March 1”. Update: you can already use it today 🚀
Same Claude Code workflow. Free until Mar 1. Then you lock the founding price.

[Wassel @wasselovski]

Good move. Nice to see microVM sandboxes becoming mainstream for coding agents. We have been running agents inside isolated microVMs since day one at Umans, on our infra. What we are exploring with envs: one ready microVM per repo, so nothing touches your laptop.
https://t.co/9kWdJZpP9D

[Wassel @wasselovski]

Love what this chart implies. Open is setting the pace now.
And yeah, it matches what I'm experimenting so far: Kimi K2.5 is Opus 4.5 level for coding, and sometimes feels a bit better.

[Wassel @wasselovski]

On a non trivial feature, parallel Claude Code agents can be a cheat code.
My flow: talk it through with Claude Code until the plan is clear, let it draft a quick spec, split into independent workstreams, spawn one agent per stream (implementation plus tests), then finish with an end to end run.

[Wassel @wasselovski]

Tested Kimi K2.5 with Claude Code. Best experience so far. Might even beat Opus 4.5.

If you want to try that setup without messing with infra, that is what Umans Code is for. Same Claude Code workflow, drop in endpoint. We stay model agnostic and serve the best OSS model for the job. If DeepSeek does not ship stronger by month end, we default to Kimi.

Founding price still available. Activates March 1. https://t.co/Sp0tKs7HgE

[Wassel @wasselovski]

Just dropped a quick demo: connect a private repo, generate a living wiki, then ask your codebase anything.
NestJS example.