A 4B model reads the note.Arxo computes the answer.

One test, six ways to take it

The same 1,100 cases. The same evaluator.

Who takes the testSizeCorrect

• Claude Opus with ArxoA frontier model reads. Arxo computes.

  —

  95.8%1,054 / 1,100

• Materia MedCalc with ArxoThe 4B model reads the note. Arxo computes.

  4B

  90.2%992 / 1,100

• Materia MedCalc standalone, calculator toolThe 4B model answers alone. A calculator does the arithmetic.

  4B

  84.9%± 2.5, three runs

• Materia MedCalc standaloneThe 4B model answers alone, no tools.

  4B

  82.2%± 0.4, three runs

• Materia MedCalc standaloneThe 0.8B model answers alone, no tools.

  0.8B

  70.9%780 / 1,100

• Qwen3.5-4B, no trainingSix examples in the prompt.

  4B

  42.8%471 / 1,100

Standalone is Materia MedCalc answering alone, without Arxo. The 4B standalone figures are means over three training runs. The 90.2% row and the 0.8B row are one training run each. With one example of the same calculator in the prompt and no training, Qwen3.5-4B answers 723 of 1,100 (65.7%) and Qwen3.5-0.8B answers 346 (31.5%).

By calculator type, Materia MedCalc with Arxo answers 709 of 720 formula cases (98.5%) and 283 of 380 scoring-scale cases (74.5%). Standalone, without tools: 91.4% and 64.7% for the 4B model, 81.8% and 50.3% for the 0.8B model. All 55 calculators are listed in the repository.

Next to published results

Systems that start from the same input, the clinical note and the question, on MedCalc-Bench Verified. The circle is the model’s size: its area is the number of parameters.

SystemSizeCorrect

• Materia MedCalc with ArxoTrained on the task · Arxo is in early access

  4B

  90.2%

• Qwen3-8B extracts inputs, calculators written by GPT-5.5Trained on the task · arXiv 2610.21477, Table II

  8B

  86.73%

• Claude Sonnet 4.6, code agent with PythonNot trained on the task · reported in arXiv 2610.21477

  —

  86.24%

• Materia MedCalc standalone, calculator toolTrained on the task · the repository, three training runs

  4B

  84.9%

• Materia MedCalc standalone, no toolsTrained on the task · the repository, three training runs

  4B

  82.2%

• Nemotron 3.5 Super VL (120B) with LoRA, direct answerTrained on the task · Fastino model card

  120B

  72.45%

• Materia MedCalc standalone, no toolsTrained on the task · the repository

  0.8B

  70.9%

• Qwen3-8B, supervised fine-tuning, direct answerTrained on the task · arXiv 2610.21477, Table II

  8B

  69.46%

• Nemotron 3.5 Super VL (120B), direct answerNot trained on the task · Fastino model card

  120B

  65.91%

• Qwen3-8B, reinforcement learning, direct answerTrained on the task · arXiv 2512.19691, Appendix E; December 2025 snapshot of Verified

  8B

  64.2%

• Qwen3-8B, one example in the promptNot trained on the task · arXiv 2610.21477, Table II

  8B

  62.73%

• GLM-4.6V, direct answerNot trained on the task · arXiv 2603.02222, Table 3

  —

  51.9%

• Qwen2.5-32B writes Python from the note aloneNot trained on the task · arXiv 2609.10728, Table 1

  32B

  44.71%

• Best direct prompting on the MedHELM leaderboardNot trained on the task · as cited in arXiv 2603.02222

  —

  34.8%

• Qwen2.5-32B, direct answer from the noteNot trained on the task · arXiv 2609.10728, Table 1; first 120 words of the note

  32B

  28.67%

On its own, the 4B model scores above a 120B model with LoRA trained on the task: 82.2% against 72.45%. The 0.8B model scores 70.9%, next to an 8B model fine-tuned on the task at 69.46%. Sizes are the ones in the model names; a dash means the name gives none.

Handed the formula and the correct input values in the prompt, a 32B model that writes Python reaches 90.53% (arXiv 2609.10728, Table 1). Materia MedCalc with Arxo reaches the same level from the note alone.

What we did not do

• Open the test early

  The test split stayed sealed until measurement.

• Train on test notes

  Training notes that also appear in the test set were removed.

• Grade ourselves

  No prompt was tuned on test cases. Every figure uses the benchmark’s evaluator.