I made a 380m open source decision model that matches its teacher*, Clef-Flash 9.4B
retained text
TLDR: matches the teacher in overall accuracy at 1/25th the size. Can run on CPU to keep precious VRAM free. In a comparison of general accuracy, diakrisis-espresso-380m achieved 3rd place, surpassed by Jev 1.14 and Google Decision 1. Main weaknesses are in policy compliance and multi-step questions. [My evaluation. No external evals yet.](https://preview.redd.it/g4e174k3gf2i1.png?width=1920&format=png&auto=webp&s=cf3540bea0fbdb66325a625c495f33b48f5407c9) Was using Jev for a bit. Wanted something private. Clef-Flash came out, and I only have 16GB VRAM. Figured hey, these are like them good ol' BERT models right, what if I just distilled Clef into a BERT model. So I did a lil diggin with Claude, decided on vlea3-fast\*, and over the last few weeks played a lil bit of wackamole and a game of find open-source data that won't get me sued (some of these open datasets had outputs from GPT for example, uh-uh not playing that game). The results are above. Not only that, but using ONXX runtime made this run on CPU to save me that sweet sweet VRAM for LLMs, including an 8-bit variant that agreed with the larger fp32 variant 412/414 times. If anyone has any suggestions for improving the model, I would greatly appreciate it! 100% open-source and open-weight, forever. Cheers! [https://huggingface.co/diakrisis/espresso-380m](https://huggingface.co/diakrisis/espresso-380m) [Git](https://github.com/CaffeinatedSoftwareLLC/diakrisis)
document fetched 2026-11-19 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xzpof4/i_made_a_380m_open_source_decision_model_that