Made a new 350M model to compete with lfm2.5 but with an open license

I liked the idea of a nano llm, but decided to actually challenge myself with developing one.

Keep in mind I developed this model, do your own research and your own benchmarks.

It has been a while since I posted on this subreddit, been busy getting better. trained, fine tuned, generated data for many llms though unreleased as they were unsatisfactory. 2.0 is not from scratch, though working on doing that too.

In the screenshots I accidentally locally saved it as fijik2.5! The model is the same one as the one uploaded on HF in bf16. My apologies.

Been working and can finally release Fijik 2.0 350m, based off of granite 4 350M, continually pre trained on ~6B tokens with an Aug 2025 knowledg cutoff, then post trained on a custom sft corpus with mixed reasoning efforts. Also, I've included some samples of outputs from the model compared to lfm2.5. Keep in mind, you should use it with web search or similar, you can't have much knowledge at 350M parameters.

Basically, lfm2.5 is awesome truly, but I don't like the custom license, fijik uses apache-2.0, and unlike my previous model(s) I actually benchmarked it. Benchmarks are available on the HF readme!

If you have any questions feel free to ask, worked pretty hard on it and honestly, I'm pleased.

GGUF: https://huggingface.co/Pinkstack/fijik-2.0-350m-sft-GGUF (running below bf16 is not recommended, you may need to set the chat format manually in lm-studio and alike, the model does NOT use standard chatml and will not work with chatml.)

Have a good one. Once again if you have questions feel free to reach out.