NVIDIA’s new FP4-Local checkpoint re-quantizes GLM-5.3-Flash with quantization-aware distillation and compressed expert scales, cutting it to 164.6 GiB with near-baseline accuracy — but it requires a forked vLLM stack (B12X kernels, modelopt_mixed loader), and the checkpoint’s own author says 4× Spark clusters gain little from it: the quant exists to buy back KV cache on 2-node setups, a constraint four nodes don’t have.

RadixArk, NVIDIA, RedHatAI, LibertAI and local-inference-lab all ship an NVFP4 GLM-5.3-Flash that differs far less in size (195-204 GB) than in quantization scheme: W4A4 vs weight-only, what stays BF16, MTP handling, calibration, and one checkpoint whose expert weights are actually retrained — with the tensor-scale math to verify each card’s claims. A same-day appendix widens the lens to the 20 most-downloaded quants of any precision, accuracy only; a second-day appendix puts the two most heavily audited cards — canada-quant’s int4 W4A16 and NVIDIA’s NVFP4 — side by side on evidence quality and scheme mechanics, and explains why no honest head-to-head number exists yet.

TensorFold’s headline GLM gains come from EXL3-trellis kernels and a lane scheduler that are quantization- and topology-specific — upstream CUDA is 10 days old, TP2-only, and one-request-at-a-time for GLM. For an FP8 checkpoint on TP4-over-switchless-ring with DFlash2 + SparkCache, the verdict is pass-and-watch; the one transferable idea is latent FP8 KV.

The 125B Qwen3.8-Flash-Next MoE really does run on a 12GB GPU — via Strata, whose experimental SYCL port already hit 70-78 tok/s decode on an Arc Pro B70. But the bigger surprise is on the B70 itself: 35B-A3B-class MoE models already clear 40 tok/s decode and 1000 tok/s prefill on plain llama.cpp SYCL. Evidence and caveats for both paths.