# Lean

- **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:** [Mehardeep Singh](https://cerebralvalley.ai/u/mehaarzz), [Gurarpit Singh](https://cerebralvalley.ai/u/arpitzzz)
- **GitHub:** https://github.com/Mehardeep-Singh-GG/lean-meta-hackathon
- **Demo video:** https://youtu.be/DtdFJmFgbKI
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3963

Meta's open-weight Muse-Glimmer-30B is a powerful 30B reasoning and multimodal model, yet developers and students cannot run it locally: at 55.46 GiB uncompressed (15.85 GiB in 4-bit), it immediately crashes an 8 GB laptop. Disk swapping crawls at 0.15 tokens/sec (seven seconds per word), while crushing all layers below 2 bits drops multi-step tasks to 0.0%.

We built Project LEAN to run frontier 30B reasoning privately and offline on 8 GB hardware at zero cloud cost. Profiling all 52 layers showed ~20% is the model's thinking engine and ~80% is passive trivia storage. LEAN keeps that 20% core sharp at 4-bit, packs the 80% bulk five values per byte while cancelling 47% of rounding error without retraining, reuses one 20 MiB slot for zero memory spikes, and uses an 18 KB local index (0.2 ms) for exact facts. Holding all 52 layers in RAM avoids slow SSD reads for real-time speed.

At 6.08 GiB (9.12x smaller), peak memory stays at 6.87 GiB (1.13 GiB free), streaming at 19.41 tokens/sec in RAM (129.4x faster than SSD swap) and 11.26 tokens/sec over PCIe. Across 5 seeds, fact recall jumps from 25.6% to 98.4%, 6-step tool tasks hit 95.0% (vs. 0.0%), pure logic hits 94.0% (vs. 40.0%), and Python coding reaches 88.0% Pass@1 (98.7% Pass@3) and 64.4% HumanEval (beating GPTQ 60.4% and HQQ 43.2%), matching the 55.46 GiB original by 99.24%.

Everything is 100% open today: download the 6.08 GiB weights, hardware receipts, and benchmarks in our repo.

This brings datacenter-grade 30B AI to millions of students, clinics, and builders on 8 GB laptops. Private, zero-cost frontier intelligence now runs on the computer already on your desk.

## More from Meta Global AI Developer Hackathon

- [CloveaXD](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3960)
- [Andaman Khunaprapakorn](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3961)
- [GeoSAR](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3962)
- [Spaces](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3964)
- [Team 1](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3965)
- [Revel Gifts](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3966)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3963. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt