Liquid AI Releases LFM2.5 QAD Q4_0 Checkpoints, Recovering 97% of Accuracy Loss
Key Highlights
Liquid AI has released a set of QAD Q4_0 quantized checkpoints named LFM2.5, comprising four models: 230M, 350M, 1.2B-Instruct, and 2.6B. The most eye-catching figure is this: while keeping the native Q4_0 memory footprint and inference speed, it recovers 97 percent of the average BF16 accuracy loss. For anyone running models on commodity hardware, that recovery rate is the difference between a usable assistant and a frustratingly degraded one.
What Happened
QAD stands for quantization-aware distillation, which, in simple terms, is a "slimming" technique: a large model acts as the teacher and "imparts" its capabilities to a smaller model that has been compressed through quantization. In the past, quantization (such as compressing to Q4_0) saved memory and ran fast but often came with noticeable accuracy drops. LFM2.5's approach uses distillation to recover that lost accuracy, making the small model both light, fast, and not much "dumber." The result is a family of checkpoints that behave far closer to their full-precision parents than naive quantization alone would allow.
The Technical Details
Q4_0 is the most common 4-bit quantization scheme in the GGUF format, prized for its small size and broad support across virtually all local inference frameworks. The four LFM2.5 checkpoints span a range from embedded devices (230M) to desktop-class applications (2.6B), and the 1.2B-Instruct variant is further optimized for instruction following, making it better suited for direct conversation and agent scenarios. Recovering 97 percent of the accuracy loss means users can get an experience close to the original model using roughly one-quarter of the storage, a compelling proposition for devices where every gigabyte counts.
Compared with Competitors
In the on-device small-model arena, Liquid AI's rivals include Microsoft's Phi series, Google's Gemma, and Meta's smaller Llama variants. LFM2.5's differentiation lies in packaging "quantization plus distillation" into ready-to-use GGUF checkpoints, sparing developers the trouble of wrestling with quantization schemes themselves. Compared with simply releasing BF16 weights, this "already compressed and accuracy-repaired for you" delivery is far friendlier to local deployment, and it lowers the expertise barrier for teams without dedicated quantization engineers.
Industry Impact and Outlook
The most direct value of these checkpoints is enabling laptops, phones, and even Raspberry Pi-class hardware to run capable language models smoothly. They are especially useful for privacy-sensitive, network-constrained, or offline scenarios such as local assistants, industrial control terminals, and edge devices. As on-device AI becomes fiercely contested territory, Liquid AI offers developers and hardware makers a new option with models that are "lightweight but not compromised," also lowering the barrier for frontier capabilities to trickle down to the terminal, where they can run privately and indefinitely without recurring cloud bills.
The broader significance of LFM2.5 is that it narrows the gap between "good enough to ship" and "small enough to embed." As more applications move inference to the edge, including phones, cars, appliances, and industrial sensors, the economics of model size become decisive. A checkpoint that recovers 97 percent of lost accuracy while fitting in a fraction of the memory makes it far easier to justify putting a capable model directly on a device, where latency is near zero and privacy is inherent. This also lowers the cost barrier for startups that cannot afford constant cloud inference fees at scale. Liquid AI's bet is that the future of useful AI is not one giant model in a distant cluster but many specialized small models distributed across the things people already own. If that bet pays off, techniques like QAD will become standard practice rather than a research curiosity, and the phrase "runs offline on a modest device" will stop being remarkable. The 2.6B variant in particular offers a taste of what desktop-class local assistance can feel like without a discrete GPU.
What makes the LFM2.5 approach notable is its reproducibility. Quantization-aware distillation is not exotic; it is a documented recipe that other teams can adopt, which means the "small but capable" bar may rise across the entire on-device field rather than staying a Liquid AI specialty. That is good news for users, who stand to benefit from a wave of compact models that respect both memory budgets and quality expectations. The 97 percent recovery figure sets a benchmark that competitors will now be measured against.