# tabs_over_spaces

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Anagha Hambir Langhe](https://cerebralvalley.ai/u/a_lan__)
- **GitHub:** https://github.com/ana-lan/thinkABR
- **Demo video:** https://www.youtube.com/watch?v=1W7iQVMgkY0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2308

ThinkABR answers a question every team running AI agents faces: when does self-hosting an open model on a cheap cloud GPU beat paying per token? I served Meta’s open-weights Muse Glimmer 30B (4-bit, vLLM) on a single NVIDIA L4 and ran a real agent on it. The agent triages data pipeline failures over 40 graded test cases with tools for logs, code, schemas and SQL. Then I turned the measured tokens, cache hits and throughput into cost per completed task.

The answer: a tuned spot L4 beats DeepInfra once it’s about 51% busy, and it never beats the cheapest provider. For agents, the cache price matters more than the GPU. Low reasoning matched high on accuracy (39/40) with about 30% fewer tokens, and an escalate-when-stuck router acts as a safety net. Muse Spark 1.3 on the same harness solved 40/40 at about 5x the cost, so Spark is the quality ceiling and Glimmer is the cost floor.

Live dashboard: thinkabr.vercel.app. Built on Meta open weights, self-hosted on my own GPU, so please consider it for Best Build on Meta Open Weight Models.

## More from Meta Global AI Developer Hackathon

- [ZeroLatency](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2305)
- [RecordioAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2306)
- [Steven Tenorio Castillo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2307)
- [Joy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2309)
- [nef](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2310)
- [lxna_builds](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2311)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2308. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt