Z.ai leads open-model rankings with unreleased weights, while Liquid AI recovers quantization losses
The latest open-model benchmarks are led by a system whose weights remain restricted, while edge devices gain a robotic control model and a quantization method that retains almost all high-precision performance.
Artificial Intelligence··Night
Z.ai GLM-5.3
Chinese startup Z.ai's new GLM-5.3 scores 60 on the Artificial Analysis Intelligence Index, matching Kimi K3 to lead the open-model table at a lower operating cost. The company is withholding the model's weights from public release for roughly two weeks, limiting access to selected security partners. The delay stems from the model's unexpectedly strong ability to identify security vulnerabilities, suggesting that top benchmark scores for open systems will increasingly clash with safety restrictions.[1]
Liquid AI
Liquid AI released Q4_0 checkpoints of its LFM2.5 edge models, reaching up to 2.6 billion parameters. These models undergo quantization-aware distillation, a process that transfers knowledge from a high-precision teacher to a quantized student. The method reportedly recovers about 97 per cent of the accuracy normally lost when compressing weights to 4-bit precision across six evaluations.[2]
NVIDIA Cosmos
NVIDIA published post-training recipes for Cosmos 3 Edge, a 4 billion parameter model designed to fit within the onboard memory of a Jetson Thor module. The system integrates a 2 billion parameter Nemotron reasoner and succeeds in roughly 23 per cent of manipulation tasks during closed-loop evaluation. The hardware constraint forces the model to generate action chunks efficiently, achieving a rate of 1.53 seconds per chunk.[3]
