Upload model_catalog.json with huggingface_hub

model_catalog.json

@@ -1,5 +1,5 @@
 {
-  "note": "Catalog of UltraTensor spliced DeepSeek-V4-Coder GGUFs. Built with ultratensor/gguf_keep.py from the public BatiAI DeepSeek-V4-Pro Q3_K_M GGUF. Redistribution: verify the upstream DeepSeek model license before publishing to third parties.",
+  "note": "Catalog of UltraTensor spliced DeepSeek-V4-Coder GGUFs. Built with ultratensor/gguf_keep.py from the public BatiAI DeepSeek-V4-Pro Q3_K_M GGUF. Licensing: upstream deepseek-ai/DeepSeek-V4-Pro is MIT and the BatiAI GGUF inherits MIT (verified 2026-08-21); redistributable with attribution.",
   "serve_flags_common": "--no-op-offload is mandatory on every Q3_K tier (the engine's default op-offload corrupts CPU-resident Q3_K dequants and crashes at >=36 prompt tokens).",
   "models": [
     {
@@ -71,6 +71,16 @@
       "desc": "IQ2_XS requant of keep8u. Fastest per GiB on the ladder.",
       "serve": "Same as keep16u-iq2xs.",
       "tiers": ["gpu"]
+    },
+    {
+      "id": "keep64-iq2xs",
+      "file": "DeepSeek-V4-Coder-keep64-iq2xxs.gguf",
+      "bytes": 102376842624,
+      "quant": "IQ2_XS",
+      "bpw": 2.36,
+      "desc": "IQ2_XS requant of keep64 (159858.29 -> 97628.97 MiB, 2.36 BPW, chunked expert-slice requant). MIXED-E: hash layers 0-2 keep 384 experts, dense layers 3+ keep 64 behind a full-width gate. Requires the UltraTensor llama.cpp fork (>= b4c34be): stock llama.cpp rejects the file via check_tensor_dims. Dense-layer selections beyond the kept stack route to fallback expert 0 (the hash-layer remap convention); the numpy reference silently drops them instead - both are documented approximations of the 2.3% out-of-stack routing tail.",
+      "serve": "CPU-only on 30+ GB RAM (mmap stream): llama-cli -m <file> -n 8 -t 8 -c 128. Load takes ~20 min over 1GbE NFS.",
+      "tiers": ["cpu"]
     }
   ]
 }