kd-web4-s0), Q4_K_MThe production router of the AI-231-ME3 voice agent as a ready-to-serve GGUF, so that nobody has to merge the adapter and
convert it. It is the LoRA adapter martinnavs/granite-4.0-h-350m-router-kd-web4-s0
merged into ibm-granite/granite-4.0-h-350m and quantised to Q4_K_M. Read that model card first: it says what the router
does, how it was trained and tested, and where it is wrong. This is an adapter-derived router, not a general assistant: it
writes one tool-call line per request (NAME|arg|arg) or escalate|<reason>, only works with the exact system message and
grammar the voice agent sends, and behaves unpredictably outside that setup.
| File | router-Q4_K_M.gguf, 222,662,144 bytes |
| SHA-256 | 51d2dd3c599db9225f514ed16db82ec9146c55caf8f4c74c28377470e9581e3e |
| Made by | merge of the adapter into the bf16 base, convert_hf_to_gguf.py to f16, llama-quantize to Q4_K_M; llama.cpp commit 8ca4060b7e30e76baaa5670ee1ec1b760b7f8552 (2026-10-28) |
| Served with | llama-server --jinja -c 8192 -np 1 -fa on -t 3 -tb 3 -m router-Q4_K_M.gguf (CPU, 3 threads; about 0.2 s per call on a laptop) |
| Checked | served by the prebuilt llama.cpp release b12345 (Linux x64 CPU): the same router lines as the production server on 12 of 12 test phrases |
| Code, data, training | https://github.com/Martinnavs/ME3-LoRA-training |
| The voice agent that uses it | https://github.com/Martinnavs/AI-231-ME3 (make setup-router downloads this file and a llama.cpp release) |
Licence: Apache 2.0, as the Granite base model; the quantisation changes nothing about the terms.