Research

We publish what we learn, including what didn’t work.

Our research asks practical questions about money and the tools people use to understand it. We share the methods, the data and the trade-offs, so others can check our work and build on it.

Finsight: adapting an open model for financial reasoning

October 2026 · Models released under Apache 2.0

Read the full write-up

Can exposing an open base model to financial filings, news and finance instructions improve its financial reasoning, within a small budget? We adapted Qwen3-14B-Base in two stages on a single AWS machine with eight NVIDIA A100 GPUs, in about a day: continued pretraining on roughly 330 million tokens of financial text, then fine-tuning on 191,428 finance instructions. Test sets were protected with 13-gram decontamination.

Numerical reasoning improved sharply. FinQA accuracy rose from 0.019 to 0.126, and multi-turn ConvFinQA from 0.148 to 0.203. But the average across six financial benchmarks fell from 0.441 to 0.399, mostly in sentiment analysis and named-entity recognition. Fine-tuning recovered part of what continued pretraining disturbed, without fully restoring it.

The final model runs in 4-bit on a 16 GB GPU, using about 9 GB of memory, so anyone can test it.

  • Finsight-Qwen3-14B-CPT-SFT

    Final model: continued pretraining on financial text, then fine-tuning on finance instructions. On Hugging Face.

  • Finsight-Qwen3-14B-CPT

    Intermediate checkpoint after continued pretraining only, for comparison and further research. On Hugging Face.

Data analysis

Built from public sources, such as FFIEC call reports and the Bureau of Labor Statistics’ consumer price index, with methods explained in each piece.

Using our research

You’re welcome to cite our analyses with a link, and to use the Finsight models under the Apache 2.0 license. If you spot an error or want to discuss the work, write to press@maxint.com.

We also publish the libraries we build on. See open source.