TinySwallow 1.5B is a family of compact Japanese language models from Sakana AI and the Swallow Team. It includes a base checkpoint for text continuation and TinySwallow-1.5B-Instruct for conversation and instruction following. This card links to Instruct because it is the practical starting point for someone building a local assistant rather than continuing the training of a base model.
Why TinySwallow was selected
On July 13, 2026, TinySwallow-1.5B ranked first by monthly downloads among models in Sakana AIโs official Hugging Face profile, with the Instruct version in second place. Download counts change daily and do not measure quality, but they provide a verifiable answer to which open model currently attracts the most use. AIDive is therefore publishing a card for a specific neural network rather than a general Sakana AI company article.
Base or Instruct
Base was further pretrained on Japanese text and is intended for research, continued training, and text-completion work. Instruct was tuned for commands and Japanese conversational use. Chat, assistant prototypes, and local APIs should normally start with Instruct. Base may be more appropriate for custom fine-tuning where a pre-existing assistant personality or chat behavior would get in the way.
TAID and the Japanese focus
TinySwallow was created with TAID, or Temporally Adaptive Interpolated Distillation. Qwen2.5-32B-Instruct served as the teacher and Qwen2.5-1.5B-Instruct as the student. Distillation attempts to transfer useful behavior from a larger model into a smaller one that costs less to run. The model card officially lists Japanese as the language. Its Qwen foundation may produce output in English or other languages, but Sakana AI does not promise equal quality outside Japanese.
Transformers, vLLM, and 3.09 GB of weights
The main BF16 Instruct weight file is about 3.09 GB. It loads through Transformers, while the official profile also gives a vLLM example for serving an OpenAI-compatible API. Third-party quantized builds exist but should be evaluated separately. The configuration lists up to 32,768 context positions. That is a technical window limit, not a guarantee of accurate work across a long document: a compact model can lose details, hallucinate, and struggle with complex reasoning.
Licensing and the limits of an experimental model
The page carries an Apache 2.0 label, but the model derives from Qwen and was also trained on Gemma data. Sakana AI explicitly requires users to follow both the Apache terms and the Gemma terms and prohibited-use policy. Commercial use is permitted only when both sets of requirements are met. The developer calls TinySwallow an experimental prototype for research and development, so factual accuracy, safety, and fitness for a production task still require independent testing.
