Finsider Labs
All leaderboards

Open-weight financial reasoning

A selected open-weight view of the same Financial Reasoning evaluation, not a separate experiment.

AIMultiple / November 2026 / published research

Evaluation setup

AIMultiple evaluated the 238-question hard subset of FinanceReasoning (Tang et al.) with chain-of-thought prompting, model-based answer extraction, and a 0.2% relative numerical tolerance. Models added from September through November 2026 ran through OpenRouter, while the publisher notes that Opus 5.5 ran in a different execution setup. These are single publisher-reported runs, not Finsider runs, and repeated-run reliability has not been established.

Reading this comparison

Ranks compare displayed scores only within this named evaluation and selected model roster. Ties share a rank. Different benchmarks are not combined into an overall ranking. A model omitted for lack of a published result is not assigned a zero. A newer release is not automatically better on every task.

Boundaries

A benchmark answer score does not establish a model's readiness to deliver a financial diligence engagement. It does not measure CPA sign-off, evidence retention, access controls, or performance on your transaction documents.

Provenance

AIMultiple results compiled November 13, 2026, from the publisher's November 10 update. Model ML results compiled September 14, 2026. Evaluations were performed by the named publishers, not Finsider. CSV downloads record model versions, configurations, sources, and any arithmetic derivation. No comparable result for Finsider's deterministic math engine is available in these sources, so the engine is not ranked.

Performance Comparison

Accuracy (%)

1

GLM-5.3

89.50

2

Kimi K3

88.24

3

DeepSeek V4 Pro 0813

87.39

4

GLM-5.2

86.13

5

gpt-oss-120b

81.09

6

Qwen3-235B-A22B-Thinking-2507

75.63

7

Llama 4 Maverick

75.21

8

DeepSeek-R1-0528

62.18

050100

Selected models with downloadable weights and a reported score in this evaluation. Model-specific licenses apply. Results belong to these exact versions, and a successor does not inherit its predecessor's score. The publisher evaluated hosted endpoints, not Finsider self-hosted deployments.

Rank uses displayed scores, and ties share rank. Bars use a 0-100 scale. No confidence intervals are available in the cited results, so small differences should not be interpreted as statistically established superiority.

Source: AIMultiple / November 2026