Card: the sidecar now also serves FP8-table checkpoints; nvidia's NVFP4 booted on one card with it

README.md

@@ -16,10 +16,16 @@ the CPU-offload worker holds it in BF16 (95.4 GB). This repo ships the same tabl
 serves them memory-mapped straight from disk. Host RAM cost becomes page cache only,
 reclaimable under pressure.
 
-Built from the original BF16 tables, so it works with any checkpoint of this model that keeps
-them: [our mixed NVFP4/FP8 build](https://huggingface.co/primitive-ai/Qwen3.8-Flash-Next-mixed-NVFP4-FP8),
-[our plain NVFP4](https://huggingface.co/primitive-ai/Qwen3.8-Flash-Next-NVFP4), or the original
-model. It does not apply to checkpoints that re-quantized the tables themselves.
+Built from the original BF16 tables, so it works with any checkpoint of this model:
+[our mixed NVFP4/FP8 build](https://huggingface.co/primitive-ai/Qwen3.8-Flash-Next-mixed-NVFP4-FP8),
+[our plain NVFP4](https://huggingface.co/primitive-ai/Qwen3.8-Flash-Next-NVFP4), the original model,
+and, since 2026-09-06, checkpoints that ship their own table in FP8. The overlay stubs the table
+parameter and serves gathers from the sidecar, so the checkpoint's own table tensors are dropped
+before they load. That turns the table format into a non-issue:
+[nvidia's NVFP4 build](https://huggingface.co/nvidia/Qwen3.8-Flash-Next-NVFP4), whose card documents
+an 8-GPU deployment because its FP8 table cannot use vLLM's BF16-only host offload, boots on one
+96 GB card this way (234 s, 88,828 MiB, 348k-token KV pool) and scores 92.1 knowledge / 78.3
+tool-calling under our protocol. Its 53.7 GB table-and-MTP shard can be skipped at download time.
 
 ## Serve