Our next generation of small language models is here.
July 18th, 2026
Today, we're releasing a preview of TFM-2, the next generation of models from Astar Labs.
TFM-2 is a larger model than anything we've shipped before: 375M parameters, a 32k vocabulary, and a cleaner training stack that got us here in roughly a quarter of the wall-clock time of TFM-1.6. This is an early checkpoint, not the finished product. The full-scale TFM-2 is still in training, and will follow after this preview window.
TFM-2 (Preview) was trained in 2.5 days on a single NVIDIA RTX 5090, about 4x faster than TFM-1.6, mostly thanks to optimizations in the training flow, not less data or a smaller ambition.
One model is available right now:
tfm-2-0713-preview) - early conversational checkpoint of the next-gen modelTFM-1.5 and TFM-1.6 have been retired from the website to make room for it.
All scores below are 0-shot. Comparisons are against two models in a similar size class: SmolLM2-360M-Instruct and Qwen2.5-0.5B-Instruct.
| Model/Benchmark | TFM-2 (Preview) | SmolLM2-360M-Instruct | Qwen2.5-0.5B-Instruct |
|---|---|---|---|
| HellaSwag | 31.70 | 52.10 | 48.00 |
| PIQA | 60.30 | 70.80 | 67.20 |
| WinoGrande | 52.00 | 52.50 | 54.10 |
| OpenBookQA | 28.40 | 37.40 | 37.40 |
These numbers are honest, and they are not the ceiling. TFM-2 (Preview) is severely undertrained right now, with a lot of performance still left on the table. The gap to similarly sized models is real, and closing it is exactly what the full-scale training run is for.
This preview is here so you can poke at the architecture, the behaviour, and the direction early, not because the model is done. When the full TFM-2 finishes training, it will ship as an open-weight release.
You can chat with TFM-2 (Preview) right here on the Astar Labs website.
TFM is free to use. It is not affiliated with any company, research institution, or commercial entity. It's just a project. A very personal one.
Parts of this article were written by TFM-2 (Preview).