Under the Hood

What it is: A set of small, compressed AI models from Liquid AI, designed to run on everyday devices like laptops, phones, and Raspberry Pi boards.

What's new: Most compressed models lose meaningful accuracy in the process. Liquid AI's approach recovers nearly all of it while keeping the same small file size and fast speed as a standard compressed model.

How it works: The technique is called Quantization-Aware Distillation (QAD). Quantization (shrinking a model by reducing the numerical precision of its internal values, the way you'd compress a photo to a smaller file) typically costs accuracy. QAD counters this by running a full-precision "teacher" model alongside a compressed "student" model during training. The student learns to match the teacher's outputs precisely, rather than just being squashed down after the fact. The result is a model that fits in the same memory as a standard compressed file but behaves closer to the full-size original.

What it can't do: According to Liquid AI's own benchmarks — covering reasoning, instruction-following, and tool use across sets including GPQA Diamond (a graduate-level science question set) and MMLU-Pro — the QAD models recover around 97% of the accuracy lost through standard compression. Frontier-scale models were not tested; all four models here are small, ranging from 230 million to 2.6 billion parameters (the numerical values that define a model's behaviour).

Who should care: Developers and hobbyists who want a capable AI model running locally, without cloud costs or privacy trade-offs.

Try it / read it: Hugging Face — LFM2.5 QAD

What it is: A scheduling system that decides which AI jobs run on which GPUs, and in what order, to get more useful work out of the same hardware.

What's new: Most GPU clusters use FIFO scheduling (first in, first out — jobs run in the order they arrive, regardless of priority or size). This system replaces that with a constraint-aware allocator that makes each allocation decision based on job priority, duration, and demand shape.

How it works: The allocator treats each scheduling decision as a grid: every GPU, across every time slot, assigned to a job or left empty. It distinguishes between two incompatible job types. Batch jobs — training runs, bulk inference, model compression — need a fixed block of GPUs held continuously until they finish. Real-time inference is elastic: it scales up and down as user traffic changes. FIFO handles this by reserving each real-time application's peak GPU count for the full day, leaving those GPUs idle overnight. The new allocator releases GPUs during low-traffic periods and reclaims them before the next peak, freeing that capacity for batch jobs in the gap.

What it can't do: The gains are measured on seven internal benchmark scenarios. Real-world workloads vary, and the authors do not claim the 33-point utilization improvement generalises to every cluster configuration.

Who should care: Engineering teams running mixed AI workloads on shared GPU infrastructure — the scheduling gains are meaningless on a lightly loaded cluster with plenty of spare capacity.

Read it: Hugging Face Blog