Samsung compresses large models with LittleBit

Samsung Research has detailed LittleBit, a compression technique designed to make large models dramatically smaller. In tests described in a NeurIPS 2025 poster, the team reduced Llama2-13B to under 0.9GB and Llama2-70B to under 2GB by combining low-rank factorization with binary latent factors. The system targets roughly 0.1 bits per weight, delivering a nearly 31x memory cut while using multi-scale compensation to recover accuracy lost to binarization.

Benchmarks cited by Samsung say LittleBit on Llama2-7B at 0.1 bits per weight beats rival quantization methods running at 0.7 bits per weight, and the compressed format can unlock up to an 11.6x inference speedup over FP16. Samsung has released code publicly and described a LittleBit-2 successor tied to ICML 2026, but has not said whether the technology will appear in Galaxy devices.

The work arrives amid wider efforts to reduce AI memory pressure. PrismML says Bonsai 27B compresses Qwen3.6-27B from about 54GB to as little as 3.9GB, while Google’s TurboQuant compresses the key-value cache to roughly 3 bits per value and Microsoft’s BitNet pursues native 1-bit training. The common goal is to make capable models practical on phones, laptops and constrained hardware rather than relying only on larger chips.