Upload README.md with huggingface_hub

README.md

@@ -14,7 +14,12 @@ GGUF (697.3 GB, 17 shards), built with `ultratensor/gguf_keep.py`. A splice
 keeps only the expert subset a domain actually uses:
 
 - **keep64** — 64 dense experts / 384 hash experts (156.1 GiB). Holds 97.66 %
-  of code routing mass on the first dense layer.
+  of code routing mass on the first dense layer. MIXED-E: dense layers keep
+  64 of 384 experts behind a full-width gate, so stock llama.cpp refuses to
+  load it (check_tensor_dims). Requires the UltraTensor fork
+  (`ultratensor-factored`, commit b4c34be) which derives per-layer expert
+  counts from the stored tensors and routes out-of-range dense selections
+  to fallback expert 0 (same convention as the hash-layer tid2eid remap).
 - **keep16u / keep12u / keep8u** — uniform 16/12/8-expert ladder splices
   (39.3 / 32.2 / 25.0 GiB).
 - **IQ2_XS requants** of the ladder (24.7 / 20.3 / 16.0 GiB, 2.36-2.38 BPW)
@@ -22,11 +27,16 @@ keeps only the expert subset a domain actually uses:
 
 All files are single-file GGUFs (split count patched), with the MTP metadata
 defect fixed (`nextn_predict_layers = 0`). They load natively in llama.cpp
-builds with DeepSeek-V4 support (b10424 line).
+builds with DeepSeek-V4 support (b10424 line), except keep64/keep64-iq2xs
+which require the UltraTensor fork (>= b4c34be) for mixed-E loading.
 
-**Licensing**: these files are derived from DeepSeek-V4-Pro weights. Verify
-the upstream DeepSeek model license and the BatiAI distribution terms before
-redistributing further.
+keep64 smoke (2026-08-22, node2 CPU): prefill 31 tokens in 291s, decode
+8 tokens in 42.8s (0.16 t/s, NFS-bound); clean completion.
+
+**Licensing**: verified 2026-08-21 — upstream
+`deepseek-ai/DeepSeek-V4-Pro` is **MIT**, and the BatiAI GGUF distribution
+explicitly "inherits the source model license: MIT". Redistribution of these
+splices is permitted under MIT with the attribution in the model card.
 
 ## Download