https://stondo.github.io/tags/glm-5.3-flash/Recent content in Glm-5.3-Flash on Bits-EntangledHugo -- 0.162.1en-usSun, 30 Aug 2026 09:00:00 +0000https://stondo.github.io/posts/glm-5.3-flash-two-rtx-pro-6000-from-garbage-to-verified/Sun, 30 Aug 2026 09:00:00 +0000https://stondo.github.io/posts/glm-5.3-flash-two-rtx-pro-6000-from-garbage-to-verified/GLM-5.3-Flash booted on my two RTX PRO 6000 Blackwell and answered every prompt with deterministic garbage. The investigation ran from a fake kernel bug I wrote myself, through a full exoneration of the linear-attention stack, to a live per-layer bisect that cornered the real culprit in the sparse-MLA prefill path. The fix came from a completely different quantization and kernel stack, and the model now passes 261,900-token needle retrieval on my desk.