Architecture families
Six architecture families, parameterised by config. Once a family is in, adding a new model that shares it is a weights + config operation — no Zig code change required.
The six families
Source: src/arch/.
| Family | Module | Examples |
|---|---|---|
| Llama-style decoder | tomoul.arch.llama | Llama 3, Phi-4, Qwen, gpt-oss, InkubaLM |
| DeltaNet / hybrid | tomoul.arch.deltanet | Trending recurrent + attention hybrids |
| Transformer encoder | tomoul.arch.transformer_encoder | bge-m3, sentence-transformer, XLM-RoBERTa |
| Transformer decoder | tomoul.arch.transformer_decoder | Smaller generative, decoder-only |
| Encoder-decoder | tomoul.arch.encoder_decoder | Whisper, NLLB-style translation |
| LSTM | tomoul.arch.lstm | Silero VAD, lightweight legacy |
Loading a config
const cfg = try tomoul.format.config_json.load(a, "phi-4/config.json");
const weights = try tomoul.format.safetensors.open(a, "phi-4/model.safetensors");
var model = try tomoul.arch.llama.LlamaModel.fromConfig(a, cfg, weights, .{});
defer model.deinit();Adding a model
If the architecture is already supported, dropping in a new model is a
config + weights operation — no Zig code. Register it in
src/models/ to
make it discoverable via tomoul.models.<slug>.
Adding a new family
New families live under src/arch/<family>/. Each family is one Zig file
(the forward pass) plus a config struct plus a validation harness against
a reference implementation. See
CONTRIBUTING.md
in the repo for the full process.
Roadmap
- GGUF support in the format readers, alongside
safetensorsand.tl. - New families land when a model the catalog needs doesn't fit an existing one — the bar to add a family is high, the bar to add a model under an existing family is just a config + weights.
Last updated 13 May 2026Edit this page on GitHub