Release · November 2026

A finance problem, step by step.A 4B model shows its work.

Arxo Pecunia FinChain 0.1 reaches 80.2% on FinChain without tools. The 0.8B model reaches 71.5%. The benchmark’s own checker did the scoring.

The task

Solve multi-step finance problems, without tools.

Twelve domains, from financial statements and ratios to investment analysis, corporate finance, risk management and sustainable finance.

Pecunia FinChain

Correct

80.2%

2,245 of 2,800 questions.

FinChain, 2,800 questions drawn from its templates. One training run per model.Open the repository

Two sizes, with and without training

FinChain has no fixed test set: every question is drawn from a template with a random seed. We drew 2,800 questions, ten per template. The benchmark’s own checker scored the final answer, with a 5% tolerance. The circle is the model’s size: its area is the number of parameters.

  • Pecunia FinChainThe 4B model, trained on the task.

    80.2%2,245 / 2,800

  • Qwen3.5-4B, no trainingThe base model, the same prompt and limit.

    76.7%2,147 / 2,800

  • Pecunia FinChainThe 0.8B model, trained on the task.

    71.5%2,002 / 2,800

  • Qwen3.5-0.8B, no trainingThe base model, the same prompt and limit.

    45.2%1,265 / 2,800

One training run per model. Greedy decoding, thinking switched off, a limit of 2,048 new tokens; the same prompt and limit for the trained and the untrained rows.

Next to the FinChain paper

The numbers the benchmark’s authors report in their paper: Table 2, final-answer correctness, zero-shot. They are measured on the authors’ own draw of questions, so the list compares the setting and not question by question.

  • GPT-4.1Zero-shot, the authors’ draw of questions · FinChain paper

    84.7%

  • Gemini 2.5 ProZero-shot, the authors’ draw of questions · FinChain paper

    84.3%

  • Claude Sonnet 4.5Zero-shot, the authors’ draw of questions · FinChain paper

    83.3%

  • GPT-5Zero-shot, the authors’ draw of questions · FinChain paper

    82.0%

  • Pecunia FinChainTrained on the task, our draw of questions · the repository

    80.2%

  • Pecunia FinChainTrained on the task, our draw of questions · the repository

    71.5%

  • Qwen 2.5 InstructZero-shot, the authors’ draw of questions · FinChain paper

    65.4%

  • Qwen2.5-MathZero-shot, the authors’ draw of questions · FinChain paper

    62.6%

  • MathstralZero-shot, the authors’ draw of questions · FinChain paper

    54.0%

  • Fin-o1Zero-shot, the authors’ draw of questions · FinChain paper

    52.8%

  • Fin-R1Zero-shot, the authors’ draw of questions · FinChain paper

    52.8%

  • Qwen 3Zero-shot, the authors’ draw of questions · FinChain paper

    32.3%

Sizes are the ones given with the results; a dash means none is given.

By difficulty

Three levels of the benchmark. Under each: the 4B model without training, Pecunia 0.8B and the 0.8B model without training.

  • Basic1,130 questions · Qwen3.5-4B 85.6% · Pecunia 0.8B 85.2% · Qwen3.5-0.8B 65.0%

    88.5%1,000 / 1,130

  • Intermediate1,120 questions · Qwen3.5-4B 76.2% · Pecunia 0.8B 74.6% · Qwen3.5-0.8B 35.7%

    82.1%919 / 1,120

  • Advanced550 questions · Qwen3.5-4B 59.3% · Pecunia 0.8B 37.1% · Qwen3.5-0.8B 23.6%

    59.3%326 / 550

The models were trained on questions of the same kinds as the basic and intermediate templates, written in our own wording with our own numbers. The advanced templates had no counterpart in training: there training gives the 4B model no gain, and the 0.8B model stays well below the untrained 4B model.

The 2,800 questions come from 280 of the benchmark’s 290 templates. The other ten templates fail to generate a question in the benchmark’s code as published.

What we did not do

  • Train on the test

    No test question was used in training. Of the 2,800 test questions none appears in the training set, word for word or with the numbers masked.

  • Tune the prompt

    The prompt was not tuned on the test questions.

  • Grade ourselves

    The final answer is read and compared by the benchmark’s checker.

Scope

  • The question kinds of FinChain

    The model is trained for the question kinds of FinChain. On other finance questions it is untested.

  • Numbers stated in the question

    It computes from the numbers stated in the question. It does not read financial statements, filings or tables, and it has no access to market data.

  • Plain arithmetic

    It works in the plain arithmetic of the benchmark: no day-count conventions, tax rules or currency handling beyond what a question states.

  • A research release

    It is a research release and not financial advice.

How it was taught

Kept.

What is public

The models are on Hugging Face; the repository holds every question and how each answer was scored. Arxo Pecunia follows the Arxo releases on MedCalc-Bench, EHRSQL and ChemIQ.

Arxo-Pecunia-4B-FinChain-0.1
Hugging Face
Arxo-Pecunia-0.8B-FinChain-0.1
Hugging Face
The 2,800 questions: template, level, seed and text
Public.
The checker’s 1 or 0 on every question, for all four rows
Public.
Checksums of the files and of the adapter weights
Public.

Check it yourself

To check the scores, draw the questions with the code of FinChain at commit 9bd2942, seeds 7000000 to 7000009 for every template, and score the final answers with its checker, ChainEval.