GLM-5.3-Flash Gets a TensorFold-Tuned 4-Bit Quant

Mia’s AI Lab has released an EXL3 4-bit quantization of Z.AI’s GLM-5.3-Flash, tuned for the TensorFold serving engine on NVIDIA DGX Spark hardware. At the same bit width, file size, and reported throughput as the TR3-4bpw baseline, the new checkpoint produces next-token probabilities closer to the full-precision model across the team’s tests.

The Apache 2.0 Hugging Face release includes a serving recipe for two DGX Spark systems. Downloading requires acceptance of the model page’s gated-access terms. All quality and performance results below come from Mia’s AI Lab and have not been independently verified.

Four bits, less drift

KL divergence measures how far a compressed model’s next-token probability distribution moves from its full-precision reference; lower values indicate a closer match. Mia reports lower point estimates on all eight test sets. The reduction ranges from 11% to 27% when measuring the quantized routed experts alone, and from 2% to 18% under TensorFold’s serving configuration, which also converts the dense weights to 4-bit precision.

Both checkpoints were scored on identical tokens, allowing a paired statistical comparison. Seven of the eight 95% confidence intervals fall entirely below zero; the chat set crosses zero, leaving that result inconclusive. On tokens where the full-precision model assigned high confidence to its top choice, the experts-only quantization selected a different token 16% to 37% less often than TR3.