file	size	sha256
`tac-1-Q5_K_M.gguf`	2.7 GB	`1c919ee74addda61bb6490195934b7c06b1a1793051e6119a2859d045e5b0634`
`tac-1-f16.gguf`	7.5 GB	`5ba9ce916778026da6a4970b41fe4f230b9a7c30c0f0437d3630724c7c94feea`

Serving

Ollama

ollama run hf.co/CodeStrux-Tech/tac-1-gguf:Q5_K_M

Ollama registers an HF-pulled model under the full hf.co/...:Q5_K_M name. To use it with the tico client (default model name tac-1), alias it once with ollama cp hf.co/CodeStrux-Tech/tac-1-gguf:Q5_K_M tac-1, or set TICO_OLLAMA_MODEL="hf.co/CodeStrux-Tech/tac-1-gguf:Q5_K_M".

llama.cpp server

llama-server -m tac-1-Q5_K_M.gguf --jinja

The server is OpenAI-compatible at /v1. Client env: OLLAMA_BASE_URL=http://localhost:8080/v1 TICO_OLLAMA_MODEL=tac-1.

Chat template warning

The chat template is ChatML with EOS id 151645 (<|im_end|>). There are 0 think references in the template. The template is byte-identical across the merged bf16, FP8, and GGUF builds. When using llama.cpp or Ollama, ensure the server applies the ChatML template correctly — the --jinja flag on llama-server enables this.

Training data attribution

Contains information from OpenStreetMap (https://www.openstreetmap.org/copyright), which is made available under the Open Database License (ODbL) 1.0. © OpenStreetMap contributors.

For full training details, architecture, evaluation, and limitations, see CodeStrux-Tech/tac-1.