Samsung's LittleBit goes viral as '13B in under 1 GB': the method is real, but it is year-old research code, perplexity rises from 5.47 to 15.92 at 0.1 bits, and it is not a local runtime

Why it matters

【Key points】

From Oct 8 an X post claiming Samsung open-sourced a way to fit a 13B LLM in under 1 GB drew about 4,300 likes and nearly 200k views; related posts reached about 575 in two days and still run at a dozen or so per hour. The work is Samsung Research's LittleBit: factorize each weight matrix into low-rank factors, binarize them to ±1, and recover accuracy with row, column and per-rank scales, reaching sub-1-bit and as low as 0.1 bits per weight. Per the paper, Llama2-13B fits under 0.9 GB.

【Verifiable】

Code: https://github.com/SamsungLabs/LittleBit

Papers: https://arxiv.org/abs/2506.13771 (NeurIPS 2025); LittleBit-2: https://arxiv.org/abs/2603.00042 (ICML 2026)

The repo dates from Nov 2025, added LittleBit-2 in May, and on Oct 9 only fixed an unpacking bug. License: CC BY-NC 4.0, non-commercial. Paper Table 1: Llama2-7B perplexity is 5.47 at full precision, 9.08 at 1 bit and 15.92 at 0.1 bits.

【Our read】

1. This is not a new release but an older paper riding the local-model wave. The real advance is academic: prior methods collapse below 0.5 bits, and factorize-then-binarize pushes the usable range to 0.1 bits.

2. Fitting a 13B model in 1 GB is not fitting 13B-level ability in 1 GB. A jump from 5.47 to 15.92 perplexity means a much weaker model, useful for studying compression limits, not a replacement for today's 4-bit local models.

3. A phone runtime is still missing: it needs quantization-aware training (5 epochs), the repo has no llama.cpp, MLX or Ollama export, and the 11.6x speedup is a kernel-level estimate. With a non-commercial license, near-term users are researchers. Watch whether mainstream inference engines adopt it.