Best Open Source Models for Financial Analysis in 2026
The open weight models with the highest financial analysis scores in November 2026 are DeepSeek V4.1 Flash (65.4% on Vals AI Finance Agent v2.1), GLM 5.3 (65.0%), MiMo V2.6 Pro (64.3%) and MiMo V2.6 Flash (63.3%). The top proprietary models score between 67.2 and 67.9%. Open weight models have the highest score of all models in two task categories, Earnings Analysis and Precedents. Mistral Large 4 is the highest scoring open weight model from a developer outside China at 61.7%, and Qwen3.8 27B is the highest scoring model that runs on a single GPU at BF16 or FP8.
On this page:
- Best open source models for financial analysis ranked
- DeepSeek V4.1 Flash, V4.1 Flash 1017 and V4 Pro 0813
- GLM 5.3 and GLM 5.3 Flash
- MiMo V2.6 Pro and MiMo V2.6 Flash
- Mistral Large 4
- Hy4 Preview
- Kimi K3.1 Pro and Kimi K3
- Muse Spark Open
- Step 5 Preview and MiniMax M3.1 Flash
- Beam 1.0, Inkling and Nemotron 3.5 Super VL
- Inkling 2 Small
- Ling 3.0 Flash Fin, Qwen3.8 Max and Qwen3.8 27B
- Open source vs open weight: license terms for commercial use
- What benchmarks exist to test AI models for financial analysis
- Agent benchmarks for finance
- Q&A benchmarks for finance
- How the Finance Agent benchmark tests financial analysis
- Hardware needed to run open source models locally
- FAQ
- Related reading
Best open source models for financial analysis ranked
The ranking below uses Finance Agent v2.1 from Vals AI, last updated 5 November 2026. The benchmark has 926 expert reviewed questions based on SEC filings of public companies, split into 27 public, 450 validation and 449 test questions. Scores are on the private test split. Each task has a two hour limit. Each model has six tools: EDGAR search, web search, an HTML page parser, a retrieval tool over fetched pages, a calculator and a price history tool. In the public Vals AI harness, the EDGAR search tool calls the SEC-API.io full text search API. A jury of three models (GPT 6.1 Sol, Gemini 3.8 Flash and Claude Sonnet 5.5) grades each answer. Grading uses partial credit, and an answer scores above zero only when it includes every figure that the answer depends on. An answer that gives two or more candidate values for one item fails. Sizes and licenses come from each model's Hugging Face repository.
| Rank among open weight models | Model | Developer | Weights released | Parameters (total / active) | License | Finance Agent v2.1 |
|---|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash | DeepSeek | 10 Sep 2026 | 552B / 8B to 16B | MIT | 65.4% |
| 2 | GLM 5.3 | Z.ai | 18 Aug 2026 | 753B | GLM 5.3 License | 65.0% |
| 3 | MiMo V2.6 Pro | Xiaomi | 22 Sep 2026 | 1.02T / 42B | MIT | 64.3% |
| 4 | MiMo V2.6 Flash | Xiaomi | 22 Sep 2026 | 309B / 15B | MIT | 63.3% |
| 5 | DeepSeek V4.1 Flash 1017 | DeepSeek | 17 Oct 2026 | Update of V4.1 Flash | MIT | 62.9% |
| 6 | Mistral Large 4 | Mistral AI | 29 Oct 2026 | 1.05T / 49B | Mistral Ecosystem Open Weight License | 61.7% |
| 7 | Hy4 Preview | Tencent | 28 Aug 2026 | 770B / 49B | Apache 2.0 | 61.6% |
| 8 | Kimi K3.1 Pro | Moonshot AI | 4 Nov 2026 | 2.8T / 104B | Kimi K3.1 Pro License | 61.5% |
| 9 | Muse Spark Open | Meta | 15 Oct 2026 | 816B / 42B | Muse Model License | 60.7% |
| 10 | GLM 5.3 Flash | Z.ai | 26 Aug 2026 | 320B / 18B | MIT | 60.6% |
| 11 | Kimi K3 | Moonshot AI | 27 Jul 2026 | 2.8T / 104B | Kimi K3 License | 60.0% |
| 12 | Step 5 Preview | StepFun | 15 Oct 2026 | 600B / 27B | StepFun Model License 1.0 | 58.9% |
| 13 | MiniMax M3.1 Flash | MiniMax | 23 Oct 2026 | 428B / 23B | MiniMax Community License | 58.8% |
| 14 | DeepSeek V4 Pro 0813 | DeepSeek | 13 Aug 2026 | 1.7T | MIT | 57.1% |
| 15 | MiniMax M3 | MiniMax | 31 May 2026 | 428B / 23B | MiniMax Community License | 54.3% |
| 16 | Beam 1.0 | Reflection AI | 16 Oct 2026 | 501B / 23B | Apache 2.0 | 51.8% |
| 17 | Inkling | Thinking Machines | 15 Jul 2026 | 975B / 41B | Apache 2.0 | 49.9% |
| 18 | Nemotron 3.5 Super VL | NVIDIA | 15 Oct 2026 | 120B / 12B | OpenMDW 1.1 | 47.7% |
| 19 | Inkling Small | Thinking Machines | Jul 2026 | 276B / 12B | Apache 2.0 | 44.8% |
On the same leaderboard, Claude Opus 5.5 scores 67.9%, Gemini 4 Argon 67.7% and Muse Wonder Full 67.6%. The best proprietary model scores 2.5 percentage points above the best open weight model. Eight of the 19 open weight models had their weights published between 15 October and 4 November 2026. Two further models on the leaderboard have announced weight releases: Inkling 2 Small at 67.5% and Hy4 at 64.7%. Both are left out of the ranking until their weights are published.
DeepSeek V4.1 Flash, V4.1 Flash 1017 and V4 Pro 0813
DeepSeek released V4.1 Flash on 10 September 2026 under the MIT license. The model has 552 billion parameters in total. It activates 8 billion parameters during prefill and 16 billion during decoding. It reads up to 1 million tokens and accepts images as input. Its reasoning effort is an integer from 1 to 100, and this setting controls the balance of cost and accuracy per request. It scores 65.4% on Finance Agent v2.1 at $0.37 per task, and has the highest open weight score in the Financial Modeling category at 46.8%. DeepSeek V4.1 Flash 1017, released on 17 October 2026 under MIT, is an update to V4.1 Flash. It scores 62.9%, and 93.0% in Earnings Analysis, the highest score of all 37 models in that category. It also has the highest open weight score on Tax Agent Bench at 44.4%, 7th of 82 models. DeepSeek V4 Pro 0813 is the 1.7 trillion parameter release and scores 57.1%.
GLM 5.3 and GLM 5.3 Flash
GLM 5.3 from Z.ai has 753 billion parameters and is released under the GLM 5.3 License. It scores 65.0% on Finance Agent v2.1, 92.4% in Earnings Analysis, and has the highest open weight scores in Adjustments at 67.5% and on the strict All Pass metric at 51.9%. GLM 5.3 Flash is a mixture of experts model under the MIT license, with 320 billion parameters and 18 billion active per token. It scores 60.6% at $0.07 per task and 4.9 minutes per task, and has the highest open weight score in Market Analysis at 67.9%. The model card describes a reasoning_effort setting with low, high and max levels and lists deployment through vLLM, SGLang and Transformers. Vals AI ran both GLM models through Fireworks AI at max reasoning effort.
MiMo V2.6 Pro and MiMo V2.6 Flash
Xiaomi released the weights of MiMo V2.6 Pro and MiMo V2.6 Flash on Hugging Face in September 2026 under the MIT license. MiMo V2.6 Pro has 1.02 trillion parameters with 42 billion active, and MiMo V2.6 Flash has 309 billion parameters with 15 billion active. Both models have a context window of 1,048,576 tokens. MiMo V2.6 Pro scores 64.3% on Finance Agent v2.1 at $0.18 per task. MiMo V2.6 Flash scores 63.3% at $0.07 per task, and 49.3% in Precedents, the highest score of all 37 models in that category. This category covers M&A transaction multiples from S-4 filings.
Mistral Large 4
Mistral AI released the weights of Mistral Large 4 on 29 October 2026, after an API preview from 6 October. The model has 1.05 trillion parameters with 49 billion active and accepts text and images. It scores 61.7% on Finance Agent v2.1 at $0.35 per task and 6.5 minutes per task, and has the highest open weight score in General Quantitative Analysis at 84.9%. The Mistral Ecosystem Open Weight License is a modified MIT license. The API preview build scores 61.4% and is a separate entry on the leaderboard.
Hy4 Preview
Hy4 Preview from Tencent has 770 billion parameters with 49 billion active and a context window of 1 million tokens. Tencent released it on 28 August 2026 under the Apache 2.0 license. It scores 61.6% on Finance Agent v2.1 and has the highest open weight score in Disclosure Analysis at 77.6%. Tencent released the final Hy4 on 4 November 2026 as an API model with image input. Hy4 scores 64.7% on Finance Agent v2.1, and Tencent announced the weights for a later date.
Kimi K3.1 Pro and Kimi K3
Kimi K3 from Moonshot AI has 2.8 trillion parameters with 104 billion active. It selects 16 of 896 experts per token and has a context window of 1,048,576 tokens. Kimi K3.1 Pro is Kimi K3 with additional post training. Moonshot released the API on 23 October 2026 and the weights on 4 November 2026, about 1.56 TB in MXFP4. Kimi K3.1 Pro scores 61.5% on Finance Agent v2.1 and has the highest open weight score in Comparables at 63.9%, above every closed source model in that category except Inkling 2 Small at 65.0%. Kimi K3 scores 60.0% on Finance Agent v2.1 and 71.3% on the Excel Modeling Benchmark v1.1, the highest open weight score on that benchmark. Kimi K3.1 Pro scores 70.0% on the Excel benchmark.
Muse Spark Open
Meta released Muse Spark Open on 15 October 2026, the open weight version of Muse Spark. It has 816 billion parameters with 42 billion active, a context window of 1 million tokens, and accepts text, images, video and audio. It scores 60.7% on Finance Agent v2.1 and has the highest open weight score in General Qualitative Analysis at 75.3%. Vals AI ran it through Together AI at $2.76 per task. The weights are released in BF16 and NVFP4.
Step 5 Preview and MiniMax M3.1 Flash
Step 5 Preview from StepFun has 600 billion parameters with 27 billion active and a context window of 1 million tokens. StepFun released the API in September and the weights on 15 October 2026. It scores 58.9% on Finance Agent v2.1 at $0.58 per task. MiniMax M3.1 Flash has 428 billion parameters with 23 billion active, the same count as MiniMax M3, and a repository of about 252 GB. It scores 58.8% at $0.16 per task. MiniMax M3, which uses MiniMax Sparse Attention, scores 54.3%. Both MiniMax models use the MiniMax Community License.
Beam 1.0, Inkling and Nemotron 3.5 Super VL
Three of the open weight models come from US developers. Beam 1.0 from Reflection AI has 501 billion parameters with 23 billion active, released on 16 October 2026 under Apache 2.0 in BF16, FP8 and NVFP4. It scores 51.8% on Finance Agent v2.1. Inkling from Thinking Machines has 975 billion parameters with 41 billion active and accepts text, images and audio. It scores 49.9%, and Inkling Small, with 276 billion parameters and 12 billion active, scores 44.8%. Both use Apache 2.0. Nemotron 3.5 Super VL from NVIDIA has 120 billion parameters with 12 billion active and is released under OpenMDW 1.1 together with training datasets and recipes. It scores 47.7% at $0.11 per task and 1.3 minutes per task.
Inkling 2 Small
Thinking Machines released Inkling 2 Small on 5 November 2026 through its Tinker platform. It has 594 billion parameters with 17 billion active. It scores 67.5% on Finance Agent v2.1, 0.4 points below Claude Opus 5.5, at $0.18 per task. Thinking Machines announced the weights under Apache 2.0 for the week after release, following external safety testing.
Ling 3.0 Flash Fin, Qwen3.8 Max and Qwen3.8 27B
Three models from the September ranking have Finance Agent v2 scores and no Finance Agent v2.1 score as of 5 November 2026. Finance Agent v2 uses a different question set and grading, so these scores are listed separately.
| Model | Developer | Parameters (total / active) | License | Finance Agent v2 |
|---|---|---|---|---|
| Ling 3.0 Flash Fin | Ant Group | 124B / 5.1B | MIT | 54.9% |
| Qwen3.8 Max | Alibaba Qwen | 2.4T / 95B | Qwen3.8 Max License | 50.6% |
| Qwen3.8 27B | Alibaba Qwen | 27B, dense | Apache 2.0 | 48.6% |
Ling 3.0 Flash Fin is Ling 3.0 Flash with continued training on financial data, developed with financial institutions. It has a context window of 256,000 tokens and costs $0.045 per task on Finance Agent v2. The general purpose Ling 3.0 Flash scores 30.3% on the same benchmark, 24.6 points lower. Ant Group released FinFIRST alongside the model, a benchmark of 123 expert authored investment research tasks. Qwen3.8 27B is a dense model with a native context of 262,144 tokens, extensible to 1 million. At 27 billion parameters, it can be served from one 80 GB GPU at BF16 or FP8.
Open source vs open weight: license terms for commercial use
All models in the ranking publish their weights. A model that publishes its weights is an open weight model; an open source model also publishes its training data and code. The license sets the terms under which a bank, fund or data vendor can deploy a model, and whether a separate agreement with the developer applies.
| License | Models | Commercial use | Conditions |
|---|---|---|---|
| MIT | DeepSeek V4.1 Flash, DeepSeek V4.1 Flash 1017, DeepSeek V4 Pro 0813, MiMo V2.6 Pro, MiMo V2.6 Flash, GLM 5.3 Flash, Ling 3.0 Flash Fin | Permitted | Keep the copyright and permission notice |
| Apache 2.0 | Hy4 Preview, Beam 1.0, Inkling, Inkling Small, Qwen3.8 27B, gpt oss 120b, gpt oss 20b | Permitted | Keep the license and notices, state changes made to the files |
| OpenMDW 1.1 | Nemotron 3.5 Super VL | Permitted | Commercial use, fine tuning and redistribution permitted |
| GLM 5.3 License | GLM 5.3 | Permitted | A company offering the model as a service with more than $10bn in revenue over 12 months passes a security review by Z.ai first |
| Mistral Ecosystem Open Weight License | Mistral Large 4 | Permitted | A separate commercial license for any company that operates a model as a service business, at any revenue; internal use is exempt |
| Kimi K3 License | Kimi K3 | Permitted | A separate agreement for model as a service businesses above $20m in revenue over 12 months; "Kimi K3" displayed in products above 100m monthly active users or $20m monthly revenue; internal use is exempt |
| Muse Model License | Muse Spark Open | Permitted | "Built with Muse Spark Open" notice; model as a service businesses above $20m in gross revenue over 12 months notify Meta and sign a separate agreement; internal use is exempt |
| StepFun Model License 1.0 | Step 5 Preview | Permitted | Attribution at scale; a separate commercial license for device manufacturers and for model as a service businesses above $20m in revenue over 12 months; internal use is exempt |
| Qwen3.8 Max License | Qwen3.8 Max | Permitted | Permission from Qwen for model as a service or AI work assistant businesses above $50m in revenue over 12 months; model name displayed above 100m monthly active users or $20m monthly revenue |
| MiniMax Community License | MiniMax M3, MiniMax M3.1 Flash | Permitted with notice | One time notice to MiniMax below $20m annual revenue, prior written authorization above it, "Built with" and the model name displayed, listed prohibited uses |
Internal research means the model runs on the firm's own infrastructure and serves only its own staff. For this use, the Kimi K3, Muse, Mistral and StepFun licenses exempt the firm from their model as a service clauses. A data vendor that gives clients access to the model through an API falls under the model as a service definitions in the GLM 5.3, Mistral, Kimi K3, Muse, StepFun and Qwen3.8 Max licenses. The Mistral license applies its clause at any revenue, the others above a revenue threshold. Six of the ten highest scoring open weight models use MIT or Apache 2.0: DeepSeek V4.1 Flash, MiMo V2.6 Pro, MiMo V2.6 Flash, DeepSeek V4.1 Flash 1017, Hy4 Preview and GLM 5.3 Flash. The license file in each Hugging Face repository contains the full legal terms.
What benchmarks exist to test AI models for financial analysis
Two types of benchmarks test AI models on finance. Agent benchmarks give the model tools, such as EDGAR search or a spreadsheet, and score the finished answer or file. Q&A benchmarks give the model a question and the source text or table, and score the answer.
Agent benchmarks for finance
| Benchmark | Maintainer | What it tests | Size |
|---|---|---|---|
| Finance Agent v2.1 | Vals AI | Analyst tasks on SEC filings with EDGAR search, web search, a calculator and price history | 926 questions |
| Excel Modeling Benchmark v1.1 | Vals AI | LBO, DCF, M&A and operating models, comps spreads and dataroom summaries built in Excel | See benchmark page |
| Tax Agent Bench | Vals AI | Research questions on US corporate tax, answered with multi step research and cited sources | 193 questions in the test set |
| Vals Finance Index | Vals AI | Weighted combination of Finance Agent, the Excel Modeling Benchmark and Tax Agent Bench | 3 benchmarks |
| Finance Agent v2 | Vals AI | Earlier version of Finance Agent v2.1, archived in November 2026 | 927 questions |
| Finance Agent v1.1 | Vals AI | Earlier version of Finance Agent with retrieval, beat or miss and trend questions | 537 questions |
| CorpFin v2 | Vals AI | Term extraction and numeric reasoning over credit agreements of 200 pages or more, archived | 858 questions |
Q&A benchmarks for finance
| Benchmark | Maintainer | What it tests | Size |
|---|---|---|---|
| FinanceBench | Patronus AI | Questions about public companies answered from their SEC filings | 10,231 questions |
| FinQA | Research paper | Multi step calculations over tables and text from earnings reports | 8,281 questions |
| ConvFinQA | Research paper | Calculations over earnings reports across several turns of a conversation | 3,892 conversations |
| TAT-QA | Research paper | Arithmetic and comparisons over tables combined with text from financial reports | 16,552 questions |
| DocFinQA | Research paper | FinQA questions with the full report as context, about 123,000 words per document | 7,437 questions |
| BizBench | Research paper | Quantitative reasoning on business and finance, including code generation | 8 tasks |
| FinBen | The FinAI | Information extraction, text analysis, question answering, forecasting, risk management and stock trading | 36 datasets, 24 tasks |
Agent benchmarks measure the full workflow: finding the filing, reading it and calculating the answer. Q&A benchmarks measure the reading and calculation steps on a given document. The ranking in this article uses Finance Agent v2.1 because its tasks use SEC filings and tools, which matches how an analyst uses a model connected to filing data.
How the Finance Agent benchmark tests financial analysis
Finance Agent v2.1 sorts its questions into nine categories. The categories are based on the work of a second or third year investment banking analyst. The leader score is the highest score any model reaches in each category.
| Category | Example task from the benchmark | Leader score |
|---|---|---|
| Earnings Analysis | Compare Rapid7's Q3 2025 actuals against prior revenue, non GAAP operating income and ARR guidance | 93.0% (DeepSeek V4.1 Flash 1017, open weight) |
| General Quantitative Analysis | Calculate the difference in days inventory outstanding between Home Depot and Lowe's for FY2024 | 87.8% |
| Disclosure Analysis | Track Boeing's segment reporting and 787 cost recovery disclosures across FY2022 to FY2024 10-K filings | 83.3% |
| General Qualitative Analysis | Compare Walmart, Costco and Target capital allocation across capex, dividends, buybacks and debt | 79.2% |
| Adjustments | Reconcile Honeywell's GAAP operating income to segment profit across annual releases | 74.7% |
| Market Analysis | Measure a stock's reaction to an announced divestiture and relate it to the stated use of proceeds | 71.0% |
| Comparables | Rank major US banks by excess CET1 ratio over their regulatory minimums | 65.0% |
| Financial Modeling | Build a DCF, LBO or accretion and dilution model from historical ratios in the filings | 57.0% |
| Precedents | Extract EV/EBITDA multiples for industrial distribution acquisitions from S-4 filings | 49.3% (MiMo V2.6 Flash, open weight) |
For retrieval, summarization and reconciliation of filing data, the category leaders score above 70%. Tasks that compare peers, build a model from the filing figures or analyze precedent transactions score between 49 and 65%. The All Pass metric requires every check in an answer to pass. Under this metric, the highest score on the leaderboard is 58.9% (Gemini 4 Argon), and the highest open weight score is 51.9% (GLM 5.3). Vals AI also reports on tool use: the highest scoring models make relatively many tool calls.
Based on these results, a finance team can use the model to retrieve and structure filing data. An analyst then reviews any multi step valuation output against the source filings.
How these models score against the closed source models on the same benchmarks is covered in Open source vs closed source AI models for financial analysis.
Hardware needed to run open source models locally
Memory for the weights equals the parameter count times the bytes per parameter. This is 2 bytes at BF16, 1 byte at FP8 and about half a byte at 4 bit quantization. The key value cache for long filings and concurrent users needs additional memory.
| Model | Weights at FP8 | Typical deployment |
|---|---|---|
| Qwen3.8 27B | about 27 GB | One 80 GB GPU at BF16 or FP8 |
| gpt oss 120b | about 61 GB as released in MXFP4 | One 80 GB GPU |
| Nemotron 3.5 Super VL | about 120 GB | Two 80 GB GPUs or one GPU of 141 GB or more |
| Ling 3.0 Flash Fin | about 124 GB | Two 80 GB GPUs or one GPU of 141 GB or more |
| MiniMax M3.1 Flash | about 252 GB as released | One node with 8 GPUs of 80 GB or more |
| Inkling Small, DeepSeek V4 Flash 0731, MiMo V2.6 Flash, GLM 5.3 Flash | 276 to 320 GB | One node with 8 GPUs of 80 GB or more |
| MiniMax M3, Beam 1.0 | 428 to 501 GB | One node with 8 GPUs of 80 GB or more |
| DeepSeek V4.1 Flash, Step 5 Preview, GLM 5.3, Hy4 Preview | 552 to 770 GB | One node with 8 GPUs of 141 GB or more |
| Muse Spark Open, Inkling | 816 to 975 GB | One node with 8 GPUs of 141 GB or more |
| MiMo V2.6 Pro, Mistral Large 4 | 1.02 to 1.05 TB | One node with 8 GPUs of 192 GB, or several nodes |
| Kimi K3.1 Pro | about 1.56 TB as released in MXFP4 | Several nodes |
| DeepSeek V4 Pro 0813, Qwen3.8 Max, Kimi K3 | 1.7 to 2.8 TB | Several nodes |
Mixture of experts models keep every expert in memory and compute only the active parameters for each token. DeepSeek V4.1 Flash has 8 to 16 billion active parameters. It generates tokens at a speed closer to a dense model of that size and needs the memory of a 552 billion parameter model. Mistral Large 4, Beam 1.0, Muse Spark Open and Inkling also ship NVFP4 checkpoints, which halve the memory of FP8. Firms can also keep the weights under their control on dedicated instances at a cloud provider, inside their own cloud account.
FAQ
What is the best open source model for financial analysis?
DeepSeek V4.1 Flash has the highest open weight score on Vals AI Finance Agent v2.1 as of 5 November 2026, at 65.4%. The top proprietary model scores 67.9%. DeepSeek V4.1 Flash is released under the MIT license, has 552 billion parameters and costs $0.37 per task on the Vals AI run. GLM 5.3, MiMo V2.6 Pro and MiMo V2.6 Flash score 65.0, 64.3 and 63.3%.
What is the best open weight model from the US or Europe?
Mistral Large 4 from Mistral AI scores 61.7% on Finance Agent v2.1, followed by Muse Spark Open from Meta at 60.7%, Beam 1.0 from Reflection AI at 51.8% and Inkling from Thinking Machines at 49.9%. Inkling 2 Small from Thinking Machines scores 67.5%, and its weights are announced under Apache 2.0.
Is there an open source model trained for finance?
Ling 3.0 Flash Fin from Ant Group is trained on financial data on top of the general purpose Ling 3.0 Flash. It is released under the MIT license with 124 billion parameters and 5.1 billion active. It scores 54.9% on Finance Agent v2, 24.6 points above the general purpose Ling 3.0 Flash.
Can an open source model run on a single GPU for financial analysis?
Qwen3.8 27B is a dense 27 billion parameter model under Apache 2.0. It fits on one 80 GB GPU and scores 48.6% on Finance Agent v2. gpt oss 120b from OpenAI also runs on one 80 GB GPU in its MXFP4 release. Nemotron 3.5 Super VL at 120 billion parameters and Ling 3.0 Flash Fin at 124 billion parameters fit on one GPU of 141 GB or more at FP8. Nemotron 3.5 Super VL scores 47.7% on Finance Agent v2.1. Larger mixture of experts models need a multi GPU node.
Are open source models good enough for financial analysis compared with GPT, Claude and Gemini?
On Finance Agent v2.1, the best open weight model scores 2.5 percentage points below the best proprietary model. 14 of the 19 open weight models score above GPT 6 Luna at 56.8% and GPT 6.1 Terra at 56.0%. In Earnings Analysis and Precedents, an open weight model has the highest score of all models. Financial Modeling shows the largest difference, with 57.0% for the best proprietary model and 46.8% for the best open weight model.
Can banks and funds use these models commercially?
Models under MIT, Apache 2.0 and OpenMDW 1.1 allow commercial use when the user keeps the license notice. These include the DeepSeek V4 family, MiMo V2.6 Pro and Flash, GLM 5.3 Flash, Hy4 Preview, Beam 1.0, the Inkling models, Nemotron 3.5 Super VL, Ling 3.0 Flash Fin and Qwen3.8 27B. Kimi K3, Kimi K3.1 Pro, Muse Spark Open, Mistral Large 4, Step 5 Preview, Qwen3.8 Max, GLM 5.3 and the MiniMax models use custom licenses with model as a service clauses, attribution clauses or notice requirements. Most of these terms apply to firms that offer the model to third parties as a service.
Why run an open weight model in place of a hosted API?
Self hosting keeps prompts, documents and outputs inside the firm's own infrastructure. This supports data residency and confidentiality requirements for material non public information, client portfolios and draft deal documents. Self hosting also fixes the model version, so a validated workflow produces the same behaviour until the firm upgrades the model. It also allows fine tuning on internal research.
How does an open source model get access to SEC filings?
The model accesses filings through tools. The SEC-API.io MCP server provides EDGAR filings, XBRL financial statements, full text search, insider trades and institutional holdings as tools. Any MCP compatible agent framework can attach these tools to a self hosted model. The EDGAR search tool in the Vals AI Finance Agent harness uses the SEC-API.io full text search API. For batch pipelines, the SEC-API.io REST APIs return the same data as JSON or extracted text.
Related reading
- Open source vs closed source AI models for financial analysis: where the gap stands across the benchmarks
Connecting a model to SEC data
- MCP server documentation: endpoint and full tool list
- How to access financial statements with ChatGPT or Claude
- How to access SEC financial data in Claude
- How to access SEC financial data in ChatGPT
- Financial Analysis Prompt Library
- API documentation
- Get a free API key
Benchmarks and model sources
- Vals AI Finance Agent v2.1
- Finance Agent v2 harness on GitHub
- Finance Agent Benchmark paper
- DeepSeek V4.1 Flash on Hugging Face
- GLM 5.3 Flash on Hugging Face
- MiMo V2.6 models on Hugging Face
- Mistral Large 4 on Hugging Face
- Hy4 Preview on Hugging Face
- Kimi K3.1 Pro on Hugging Face
- Kimi K3 on Hugging Face
- MiniMax M3.1 Flash on Hugging Face
- Beam 1.0 on Hugging Face
- Ling 3.0 Flash Fin on Hugging Face
- Qwen3.8 27B on Hugging Face
Last updated: 6 November 2026. Benchmark scores and license terms change with each model release. Check the Vals AI leaderboard and the license file in each Hugging Face repository before deployment. Educational content for general information about AI models and their licenses.