Veyra AI

Tiny English language models built for fast local inference.

Veyra AI focuses on compact, CPU-friendly language models that are easy to run, fine-tune, and experiment with. Our work is centered on small English models, function calling, Python-oriented variants, distillation, RLVR, tool use, and local AI.

The goal is simple: make capable small models that are practical for local workflows, research, and lightweight deployment.

Planned releases

• veyra-smollm2-135m-thinking — Thinking variant of veyra-smollm2-135m-instruct that uses CoT.

• veyra-smollm2-135m-instruct — Instruct model built on top of veyra-smollm2-135m-base.

• veyra-smollm2-135m-base — Extended pre-training for SmolLM2-135M.

• veyra-80m-instruct — 80M parameter instruction-tuned model.

• veyra-80m-function — Function-calling base model for downstream fine-tuning.

• veyra-80m-search — Tool-use model for search, retrieval, and result summarization.

• veyra-80m-base — Base English language model.

In production

• veyra-cosine-gated-adam-tinystories-15m veyra-adamw-tinystories-15m veyra-muon-tinystories-15m — Testing custom optimizer called CosineGatedAdam for veyra-80m-base architecture.

Current releases

• veyra-30m-base-5b-tokens — Final base model for 30m. Trained on 5B tokens and increased context length.
• veyra-30m-base-2.5b-tokens — Early base checkpoint trained on 2.5B of a planned 25B tokens. Released for transparency and experimentation; not intended for production inference.
• veyra-30m-instruct — Experimental instruction-tuned model built on top of the 5B-token base checkpoint.

License

Veyra models are planned to use the Apache 2.0 license unless otherwise noted. Please retain attribution to Veyra AI when redistributing models or releasing derivative work.