# SamsungLabs/LittleBit

> Official implementation of LittleBit (NeurIPS 2025) and LittleBit-2 (ICML 2026) — sub-1-bit LLM quantization methods that factorize weights into low-rank latent factors, binarize them, and restore magnitude with learned scales.

- Magnitude: 2.9 out of 10 — Steady
- Stars: 107 total · +12 stars measured 2026-10-08, 16:51–18:02 UTC, ≈ 15 by evening
- Star trust: star growth looks organic
- Category: Language models · Language: Python · Created: 2025-11-18 · Last push: 2026-05-06
- GitHub: https://github.com/SamsungLabs/LittleBit · Page: https://gitnova.dev/en/r/SamsungLabs/LittleBit

## Useful for

- Compress Llama 2 7B to 1.0 or 0.1 bits per weight via QAT with SmoothSign
- Enable Joint-ITQ initialization with --use_itq to improve latent geometry
- Evaluate a checkpoint on wikitext2/c4 perplexity and zeroshot tasks via eval.py

## Why it’s here

- Star-counter measurements on 2026-10-08 (UTC), 16:51–18:02: 95 → 107 stars (+12). This is the change over that interval.
- Estimated end-of-day forecast: about +15 stars, using observed gains and the previous day.
- Hacker News: “Sub-1-Bit LLM Compression via Latent Factorization” — 58 points, 4 h ago.
- Recent forks include notable developers: @alexdolbun (708 followers), @danzeeeman (263 followers).

## Star trust

Star growth looks organic.

Star-trust labels are heuristics based on the repository’s behavior, not a check of every stargazer.

## Numbers

- Forks: 16
- Issues and pull requests: 21
- Watchers: 3
- Average over the last week: 2 per day
- Usual pace: 0 per day
- Stars in the last hour (measured): 10

## Stars per day, last 30 days (oldest → newest, today is partial)

2026-09-09 … 2026-10-08: 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 12

## Hacker News

- Sub-1-Bit LLM Compression via Latent Factorization — 58 points, 8 comments: https://news.ycombinator.com/item?id=50005608

## Spotted in now

- Spotted on Hacker News

## More in this category

1. **Niko1221/Strata** — 5.7 · Peaking · Language models · C++ · +1,214 stars measured 2026-10-08, 00:01–17:58 UTC, ≈ 1,579 by evening
   A C++ local inference engine that runs the 125B MoE model Qwen3.8-Flash-Next on a regular PC with one NVIDIA GPU (12-24 GB) and 64 GB RAM, exposing an OpenAI/Anthropic-compatible API on localhost.
   Full card: https://gitnova.dev/en/r/Niko1221/Strata.md
2. **mw00/project-maya** — 4.9 · Early signal · Language models · C++ · +43 stars measured 2026-10-08, 04:56–17:58 UTC, ≈ 58 by evening
   Local engine, server and dashboard to run the 321B MoE model GLM-5.3-Flash on one or two NVIDIA GPUs, offloading experts to RAM and NVMe.
   Full card: https://gitnova.dev/en/r/mw00/project-maya.md
3. **Edge0-AI/Edge0** — 4.4 · Breakout · Language models · Python · +216 stars measured 2026-10-08, 00:03–17:59 UTC, ≈ 280 by evening
   An open-source streaming MoE inference framework: expert weights are offloaded from SSD on demand while a trained prerouter predicts routing ahead of time. Runs on Apple Silicon via MLX and ships with two ready-to-run model tiers (35B and…
   Full card: https://gitnova.dev/en/r/Edge0-AI/Edge0.md
4. **deepseek-ai/DeepGEMM** — 4.3 · Peaking · Language models · Cuda · +16 stars measured 2026-10-08, 00:01–17:59 UTC, ≈ 24 by evening
   CUDA tensor core kernel library: GEMM (FP8, FP4, BF16), fused MoE, MQA scoring for the indexer. Kernels compile at runtime via DeepJIT, no CUDA build at install time.
   Full card: https://gitnova.dev/en/r/deepseek-ai/DeepGEMM.md
5. **StayLameBro/backburner** — 4.2 · Cooling · Language models · Python · +40 stars measured 2026-10-08, 00:02–17:59 UTC, ≈ 53 by evening
   A llama.cpp fork that plugs an iPhone into a Mac over USB-C and splits the work: the Mac runs layers 1-40, the iPhone runs 41-64 on its GPU, speeding up prefill and allowing context up to 196k-229k tokens. Aimed at people running…
   Full card: https://gitnova.dev/en/r/StayLameBro/backburner.md

---

Magnitude (0–10) measures how fast and how unusually interest in a repository is growing right now. It is not a quality score.
Days are UTC. “So far today” is a fact; “expected by the end of the day” is a forecast.
Summaries and use cases are written by an LLM (DeepSeek V4.1 Flash) from the README and may be inaccurate: verify specific claims (benchmarks, speed, hardware) in the repository itself.
Data as of 2026-10-08 18:19 UTC, updated every 30 minutes.