TensorFold vs vLLM for GLM-5.3-Flash FP8: what the new engine actually is, and whether its speedups can move a 4x DGX Spark ring

Calendar October 6, 2026
Dgx-Spark, Tensorfold, Vllm, Glm, Inference, Exl3, Fp8, Speculative-Decoding
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.

Qwen3.8-Flash-Next on an Intel Arc Pro B70: what Strata, TensorFold and the new specialized engines can actually do

Calendar October 4, 2026
Llm, Intel-Arc, B70, Inference, Llama-Cpp, Qwen, Strata, Tensorfold
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.