My GLM-5.1 coding agent scored 94.3% on LiveCodeBench Lite (348/369)

5 points

a month ago

I’ve been building Univence, a custom autonomous coding agent platform powered by GLM-5.1.

We are building this to be a true Replit/Vercel competitor, but with zero vendor lock-in. You can build and develop entirely on our platform alongside our SOTA agent, but you own the code and can deploy it seamlessly to any 3rd-party host like DigitalOcean, Netlify, AWS, or your own VPS.

To prove the core agent's capability, we just ran it against the LiveCodeBench Lite dataset (Python split). Here is the breakdown over a blind 369-problem run:

    Total: 348/369 passed (94.3%)

    Easy: 138/141 passed (97.9%)

    Medium: 152/156 passed (97.4%)

    Hard: 58/72 passed (80.6%)

(Note: We achieved that 80% on Hard by engineering the agent's constraints to strictly prioritize optimal time complexities like O(n log n) over brute-force O(n^2), avoiding the Time Limit Exceeded errors that usually trip up standard wrappers).

But we aren't just building this for the tech. My co-founder is a Palestinian refugee currently living in the Gaza Strip, and we are launching this to drive immediate humanitarian impact. 100% of the profits from 11 months of the year from this platform will be donated directly to support Palestinian refugees.

The agent is this good already, but I have a roadmap of architectural ideas to make it even better. Right now, I'm looking for fast angel funding, compute sponsorships, or strategic partners to help us scale this ASAP.

    Try it out: https://univence.com

    Raw JSONL trajectory logs: https://github.com/UnivenceAI/Univence-benchmarks/tree/main/Z%20AI/GLM-5.1

    Follow our progress & proof of donations: https://x.com/UnivenceAI

I would love any feedback on the platform or the agent architecture. If you are an investor or want to support the mission, my DMs are open on X, or you can reach us at [email protected].