Best LLM for Finance (2026): Open Models for Financial Analysis You Can Run In-House

Finance work punishes the failure modes language models are known for. A summary that drops a footnote, a ratio computed from the wrong period, or a confident number with no source can cost more than the hours the model saved. That is why model choice matters less than the system around it, and why banks and investment firms increasingly run the model where the data already sits.

This guide covers open-weight models you can host yourself. We checked each model’s size, licence, context window and release on its Hugging Face card on 22 October 2026, confirmed serving support in vLLM’s supported-model list, GGUF builds for llama.cpp, and Ollama’s library, and read the finance benchmarks at their source. Vendor benchmark figures are the vendor’s own.

Quick picks: which model for which finance job

The open-weight shortlist, checked in October 2026

Two licence notes matter to financial institutions. Mistral Medium 3.5, a 128B dense model, ships under a modified MIT licence that withholds all rights from any company whose consolidated monthly revenue exceeded $20 million in the preceding month, which rules out nearly every bank. NVIDIA’s Nemotron uses its own OpenMDW-1.1 terms, which your legal team should read as it would any new licence.

Developer jurisdiction is the other filter. Some institutions’ procurement or model risk policies restrict models from particular countries regardless of where the weights run. Qwen comes from Alibaba in China; Gemma, gpt-oss, Granite and Nemotron come from US companies; Mistral is French. Settle that question before the benchmark comparison, not after.

How much GPU memory each model needs

The weights below use 2 bytes per parameter for 16-bit and 4.89 bits per weight for the Q4_K_M quantization, the same arithmetic as our LLM VRAM calculator. gpt-oss ships in MXFP4, so its figures are the real GGUF file sizes.

Long documents cost memory beyond the weights. Every token of a filing you load sits in the KV cache, and a 200-page report with its notes can occupy a large part of a 24 GB card on its own. Plan for the longest document an analyst will actually send, multiplied by the number of people using the server at once. Our VRAM-by-card guide shows how much context each card keeps beside the weights.

Serving these models with vLLM, llama.cpp or Ollama

For a shared analyst service, vLLM is the usual choice: it batches many users on one GPU and exposes an OpenAI-compatible endpoint. llama.cpp and Ollama suit a single workstation or a pilot on a laptop.

Matching the model to the finance task

Annual reports and 10-K filings that need long context

Annual reports mix narrative, tables and footnotes, and the answer to a question often depends on a note fifty pages after the table it qualifies. A long window lets the model see both. Qwen3.8-27B and the Qwen3.5 family hold 262,144 tokens natively, Gemma 4 31B holds 256K, and Nemotron 3.5 Lightning reaches 1M with only a few attention layers, which keeps its cache small.

Long-context scores are a better guide than the window size. Qwen reports 63.0 on AA-LCR for Qwen3.5-9B and 66.9 for Qwen3.5-122B-A10B; NVIDIA reports 52.0 for Nemotron 3.5 Lightning; Google reports 66.4% for Gemma 4 31B on an eight-needle retrieval test at 128K. Even so, retrieval that narrows a 300-page filing to the relevant sections usually beats pasting the whole document, both for accuracy and for cost per question.

Spreadsheets, statements and table reasoning

Most errors in financial question answering start with a table that was flattened badly: merged header cells, a units row that says “in millions”, or a negative shown in brackets. Parse tables into structured rows before the model sees them, keep the column headers and units with every chunk, and ask for answers that name the row and period used.

For exports from Excel or a data warehouse, give the model a query tool rather than a pasted sheet. A model that writes SQL or pandas against the real data, and shows the query, is easier to check than one that reads 2,000 rows from its context.

Numerical accuracy: tool calls versus pure generation

Language models predict digits as text, so multi-step arithmetic in prose is where they fail quietly. Hand the calculation to a tool. gpt-oss ships with a Python execution tool and structured outputs; Mistral’s model card demonstrates a calculator tool; Qwen, Gemma 4 and Granite 4.2 document function calling. The model then decides what to compute, and code computes it.

Agentic finance tasks remain hard for small models. NVIDIA’s own table reports scores between 9.28 and 14.02 on the banking domain of τ³-bench for Nemotron 3.5 Lightning, Qwen3.6-35B-A3B, Gemma 4 26B A4B and gpt-oss-20b. Read that as a reason to keep agents on narrow, checked workflows with human sign-off, whatever model you choose. Our tool-calling model guide compares tool use in more depth.

Earnings-call summarisation

A transcript of an hour-long call fits comfortably inside any window in this guide, so the 9B and 27B models do this job well. The useful output is structured: guidance changes with the quoted sentence, analyst questions grouped by theme, and figures checked against the press release. Ask the model to quote rather than paraphrase anything management said about guidance, and to mark statements it could not find in the text.

Credit memo drafting

Drafting a credit memo is an assembly job: spreading financials across periods, restating covenants, listing collateral, and writing the descriptive sections from sourced facts. A 27B to 35B model with retrieval over the borrower file handles that well. The rating, the probability of default and the approve-or-decline recommendation should stay with validated models and the credit officer. Our guide to credit memo agents draws that line in detail.

Multilingual filings

Groups that read filings in German, French, Japanese or Arabic need a model tested in those languages. IBM lists 12 tested languages for Granite 4.2, including German, French, Japanese, Arabic and Chinese. Qwen lists 201 languages and dialects for Qwen3.5 and reports multilingual benchmarks across 29 to 55 languages. Google says Gemma 4 supports more than 35 languages out of the box and was pre-trained on more than 140. Test on real filings, because accounting terms and number formats vary by country.

Finance-specific open models: FinGPT, Fin-R1, Palmyra-Fin and BloombergGPT

Searches for an open-source financial LLM usually lead to four names. Here is what each one is in October 2026.

These models proved a point: finance data in training improves finance tasks. But each is built on a base that newer general models have overtaken, and most teams now get further with a 2026 general model, retrieval over their own documents, and tools for arithmetic. Fin-R1 is still worth testing as a small extraction or classification model on a CPU-only server.

Reading finance benchmarks with care

Three public benchmarks come up in every finance model discussion:

• FinanceBench (Patronus AI and partners, November 2023) holds 10,231 questions about public companies with answers and evidence strings. On a 150-case sample, GPT-4-Turbo with a retrieval system answered incorrectly or refused 81% of the time. The lesson was about retrieval and document handling as much as the model.
• FinQA (2021) tests numerical reasoning over financial reports, with questions written by finance experts and gold reasoning programs. It rewards models that pick the right numbers and operations, which is exactly where tool use helps.
• FinBen (2024) bundles 36 datasets across 24 financial tasks, from extraction to forecasting and trading. Its authors found models strong at extraction and weak at forecasting and complex generation.

Treat any single score with caution. Public test sets leak into training data, vendor numbers come from the vendor’s own harness, and none of them uses your chart of accounts, your credit policy or your filings. Build a test set of 100 to 200 questions from your own documents, with answers your analysts agree on, and rerun it every time you change the model, the prompt or the retrieval settings.

Why banks and investment firms keep these models in-house

Three pressures push financial institutions toward models they host:

• Material non-public information. Broker-dealers and investment advisers must maintain written policies to prevent misuse of MNPI under Exchange Act Section 15(g) and Advisers Act Section 204A. A prompt that contains deal terms or unpublished results creates a copy of that information in whatever service processes it. Keeping inference, retrieval indexes and logs on your own systems keeps those copies behind your existing information barriers, a topic our information barriers article covers.
• Client data. Account data, credit files and transaction histories carry privacy and secrecy obligations. Self-hosting avoids adding a model provider as a new recipient of that data.
• Model risk management. SR 26-2, issued by the Federal Reserve, OCC and FDIC on 17 April 2026, replaced SR 11-7 and SR 21-8. It states that generative and agentic AI models are outside its scope and that each bank’s own risk management and governance should set the controls for them. Examiners still expect an inventory, testing and change control, and those are far easier when the model version cannot change underneath you. A self-hosted model stays pinned until your own change process approves an upgrade.

The trade-off is operations. You own GPU capacity, patching, monitoring and evaluation. For most institutions that cost is smaller than the work of documenting and defending a third-party model they cannot inspect.

How VDF AI fits a finance model stack

VDF AI is software that a bank or investment firm installs on its own servers, on-premises, in a private cloud or fully air-gapped. Any of the models above can sit behind it on vLLM or Ollama. VDF AI Router registers those local endpoints, lets you flag a domain as regulated so only approved models are considered, and in air-gap mode keeps every request on local models.

For the workflows in this guide, analysts query filings and research through permission-aware retrieval, so a user only sees passages from documents they could open in the source system, which keeps deal-team material away from public-side staff. Governed agents assemble credit memo drafts or earnings summaries, call calculation tools, and pause at a human approval step before anything leaves the team. Each turn records the user, the model version, the retrieved passages and the tool calls, in storage the institution controls. VDF AI holds no regulatory approval of its own; it runs inside the institution’s controls and model governance.

For the industry view, see the finance and banking solution, the finance agents catalogue and the finance AI use cases. A finance team can test this on its own hardware during a three-month on-premises platform trial before committing.