AI
“Samsung researchers developed a new AI compression technique called LittleBit that compressed Meta's Llama 2 13B model from 26.06 GB to just 0.84 GB, achieving approximately 31x compression.”
Plain restatementResearchers at Samsung Research published a quantization method named LittleBit that reduces the stored size of Llama2-13B from 26.06 GB in FP16 to 0.84 GB, a compression factor of roughly 31x.
This one checks out on the numbers. Samsung Research did publish a compression method called LittleBit, it was accepted at the NeurIPS 2025 conference, and the paper's own table lists Llama 2 13B going from 26.06 GB down to 0.84 GB, which the authors state as a 31x reduction. That 0.84 GB is the full model size, not a figure made to look better by leaving parts out. What the headline leaves out is the cost: that size comes from the most extreme setting in the paper, where the model's score on a standard language test gets roughly three times worse, and producing the compressed model requires a GPU training run rather than a quick conversion. The 11.6x speed figure the post mentions is a measurement of one layer on a datacenter graphics card, which the paper itself calls potential rather than proven end to end. No independent group has reproduced the results, so this rests on the authors' own reporting plus peer review. The research is also no longer new, as the paper first appeared in May 2025 and the same team has already published a follow-up.
[drifted from the evidence:] Samsung researchers [drifted from the evidence:] developed a [drifted from the evidence:] new AI compression technique called LittleBit that [drifted from the evidence:] compressed Meta's Llama 2 13B model from 26.06 GB to [drifted from the evidence:] just 0.84 GB, [drifted from the evidence:] achieving approximately 31x compression.
Researchers [added by the neutral restatement:] at Samsung Research published a [added by the neutral restatement:] quantization method named LittleBit that [added by the neutral restatement:] reduces the stored size of Llama2-13B from 26.06 GB [added by the neutral restatement:] in FP16 to 0.84 GB, [added by the neutral restatement:] a compression [added by the neutral restatement:] factor of roughly 31x.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- LittleBit is a real technique described in a real paper, accepted at NeurIPS 2025, with an official code release from Samsung's GitHub organization.
- The authors are Samsung Research employees, so "Samsung researchers developed" is accurate.
- The exact figures in the claim match the paper. Table 3 gives Llama2-13B as 26.06 GB in FP16 and 0.84 GB under LittleBit, stated in the paper as 31.02×. The claim's "approximately 31x" is the paper's own number.
- The 0.84 GB is the full model footprint, including the embedding and output head that stay at FP16. The compression ratio is not manufactured by excluding uncompressed parts.
- The post's statement that smaller compressed models perform worse than full-size versions is correct for this paper.
- The post's handling of the 11.6x figure as "a specific GPU computation benchmark" is a fair description. The paper's measurement is a kernel-level benchmark on one MLP layer on an A100, not end-to-end throughput.
- The arXiv link given as the source is the correct paper.
- The 0.84 GB result is the 0.1 BPW setting, which is where the quality cost is largest. On WikiText-2, Llama2-13B perplexity goes from 4.884 at FP16 to roughly 15.09 at the 0.1 BPW setting, which is about a threefold increase in a measure where lower is better. The post says accuracy "can" be reduced and that the smallest models "performed worse," which is directionally honest but leaves a reader free to imagine a small penalty next to a 31x saving. The size number and the quality number come from the same row of the same table and belong together.
- LittleBit is quantization-aware training. Producing a 0.84 GB Llama2-13B requires a GPU training run, it is not a conversion a user applies to a downloaded model in minutes. The post's framing around running models "locally on everyday devices" skips the cost of making the compressed model in the first place.
- The word "new." The preprint first appeared on 30 May 2025 and the paper was presented at NeurIPS 2025. By the post's October 2026 date a follow-up, LittleBit-2, had already been accepted at ICML 2026.
- Whether the result has been independently reproduced. No third-party replication, audit, or refutation of the 0.84 GB footprint or the perplexity figures was found. Every substantive source traces to the authors or to the venue that accepted their paper, so the evidence is credible-by-peer-review rather than independently confirmed.
- Whether a 0.84 GB Llama2-13B actually runs usefully on phone or laptop hardware. The speed evidence is a kernel benchmark on a datacenter A100 GPU, and the paper itself calls the 11.6x figure potential. No end-to-end on-device measurement was located.
- Whether a heavily degraded 0.1 BPW model is useful for the practical purposes the post implies. Perplexity is a next-token prediction measure, and the paper's commonsense reasoning results at the lowest bit-widths were not fully resolved here.
- Whether the memory figure holds for models other than those tested. The paper covers models from 1.3B to 32B parameters, and the 31x ratio is specific to Llama2-13B.
The paper exists, is authored by four Samsung Research employees using samsung.com addresses, and was accepted at NeurIPS 2025. The abstract states that the method targets quantization rates as low as 0.1 bits per weight, "achieving a memory reduction of approximately 31×, which effectively compresses Llama2-13B to under 0.9 GB," and that code is available on GitHub. The specific figures in the claim come from the paper's Table 3. That table lists Llama2-13B at 26.06 GB in FP16, falling through intermediate settings to 0.84 GB (31.02×). Samsung's own blog repeats the same pairing: Llama2-7B goes from 13.49 GB to 0.79 GB at 0.3 BPW and 0.63 GB at 0.1 BPW, and Llama2-13B at FP16 26.06 GB "compresses to just 0.84 GB." The 0.84 GB figure is the whole-model footprint, not a transformer-blocks-only number. The paper explains the arithmetic gap directly: "The gap between the effective BPW and the compression factor arises because quantization is applied only to the Transformer blocks, while" the embedding layer and lm_head remain in FP16. An aggregator extraction of the same page records that BPW is computed over the compressed transformer blocks while embedding and lm_head stay FP16. On the quality cost at that exact setting, the paper reports WikiText-2 perplexity for Llama2-13B of 4.884 at FP16 and approximately 15.09 at 0.1 BPW, with 10.48 at 0.3 BPW. Column alignment is corroborated by two independent statements from the authors: Llama2-7B scores 15.92 perplexity at 0.1 BPW, and "For Llama2-13B at 0.8 BPW, LITTLEBIT records a PPL of 8.52, whereas STBLLM records 11.90." The method is quantization-aware training, not a one-shot post-training conversion. The paper notes that post-training methods such as GPTQ and AWQ compress to roughly 4 bits, and that going further typically requires quantization-aware training to maintain performance. On the speedup the post mentions: the kernel benchmark is on an NVIDIA A100 for a single Llama2-70B MLP layer of size 8,192 × 28,672, comparing an FP16 GEMM baseline against OneBit and the LittleBit kernel at BPW configurations from 1.0 down to 0.1, with speedups that "inversely correlate with bit-width, peaking at 11.6× over an" FP16 baseline. The paper frames this as potential rather than measured end to end: "unlocking a potential 11.6× inference speedup relative to FP16."
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Did Samsung actually create LittleBit, and is the 31x compression figure real?
Yes. LittleBit is a real technique from Samsung Research, accepted at NeurIPS 2025, and the paper's own Table 3 lists Llama2-13B shrinking from 26.06 GB to 0.84 GB, which the authors call a 31.02x reduction. The claim's numbers match the paper directly.
Is the 0.84 GB figure the whole model, or does it leave parts out?
It is the full model footprint, including the embedding and output head layers, which stay in FP16. The compression ratio is not inflated by excluding uncompressed parts of the model.
What does the claim leave out about the cost of that compression?
The 0.84 GB size comes from the most extreme setting in the paper, where WikiText-2 perplexity for Llama2-13B worsens from 4.884 at full precision to roughly 15.09, about a threefold drop in quality. Producing this compressed model also requires a GPU training run, not a quick conversion a user could do at home.
Can a device really run this compressed model 11.6x faster?
That 11.6x figure comes from a kernel-level benchmark on a single MLP layer using a datacenter A100 GPU, not a full end-to-end test on everyday devices. The paper itself describes this as a potential speedup rather than one it measured in practice.
Has anyone outside Samsung confirmed these results?
The investigation did not find any independent reproduction, audit, or refutation of the 0.84 GB figure or the perplexity numbers. All evidence traces back to the paper's authors or the peer review process at NeurIPS, so it is credible through peer review but not independently verified.