GGUF quantizations of TIM-360M-base for llama.cpp and downstream runtimes (Ollama, LM Studio, llamafile). This is the base model — a from-scratch 361.8M-parameter decoder-only transformer. It is not instruction-tuned; for chat use tim-360m-instruct.

Converted from the Qwen3ForCausalLM export (arch: qwen3, byte-level BPE, Llama-3-family pre-tokenizer).

At this scale everything runs comfortably on CPU.

Ollama (Modelfile):

This is a base completion model — give it a prefix to continue, not a chat prompt.

Apache-2.0. See the source model card for architecture, training data, and evaluation. Pretraining data (all named, public): FineWeb-Edu (ODC-BY), DCLM-Baseline (CC-BY-4.0), Stack-Edu, FineMath (ODC-BY).