# qwen-dianjin: a hub of seven financial AI projects, and a matrix of what is downloadable > Qwen DianJin is Alibaba Cloud's open source collection for financial AI, and it is not a library. It is seven sub-project directories plus a README table with five columns that tells you which of them have code, models, data and papers, and which have only some of those. **aliyun/qwen-dianjin** — Qwen DianJin: LLMs for the Financial Industry by Alibaba Cloud（通义点金：阿里云金融大模型） - Repository: https://github.com/aliyun/qwen-dianjin - Website: https://tongyi.aliyun.com/dianjin - Stars: 625 · Forks: 68 - Language: Python - License: MIT - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/aliyun-qwen-dianjin ## Sixteen root entries, seven of which are projects The repository root contains `.gitignore`, `CODE_OF_CONDUCT.md`, `CONTRIBUTING.md`, a `LICENSE`, a `LICENSES/` directory, a `NOTICE`, two READMEs in English and Chinese, an `images/` directory, and seven project directories. Those seven are `DianJin-CSC`, `DianJin-OCR-R1`, `DianJin-PRM`, `DianJin-R1`, `DianJin-RED`, `DianJin-SKILLS` and `DianJin-TIR`. Each one has its own README linked from the table in the root document, which is what makes this a hub rather than a monorepo with a shared build. There is no root manifest and no root package to install. The recorded primary language is Python, which reflects the contents of the sub-projects rather than anything at the top level. The README describes the repository as serving as the open source hub for the financial AI research, with released code, models and data summarised in one table. That framing explains the whole document: it is an index, and the projects are the artefacts. ## The directory names do not match the project names in the table Searching the repository for a name you read in a paper will not find the code. Fin-PRM lives in `DianJin-PRM`. FinMCP-Bench lives in `DianJin-TIR`. The customer support conversation work, titled as evaluating, synthesizing and enhancing for customer support conversation, lives in `DianJin-CSC`. The table's own Project column uses the paper names while the Code column uses the directory names, and the two only sometimes agree. There is a plausible reason for the mismatch in FinMCP-Bench. Its technical report lives inside the DianJin-TIR directory with a technical report filename, so TIR appears to be the umbrella name for that line of work rather than the benchmark's own short name. The practical consequence is small but real: a first-time visitor who tries the project name from an arXiv listing and finds nothing will conclude the code is not there, when it is under a different name one row away. ## A five-column availability matrix, and most rows are sparse The table has five columns: project, code, ModelScope, HuggingFace and paper. A dash means the artefact is not there. Reading the rows tells you exactly what you can download today. Three projects are complete. Fin-PRM has code, a ModelScope entry, a HuggingFace entry and an IJCAI 2026 paper. DianJin-OCR-R1 has all four with its arXiv identifier. CSC has all four, and its two hub entries are datasets rather than models. Two are code only. DianJin-SKILLS and DianJin-RED have a linked README and dashes everywhere else, which is what you would expect from skill libraries and benchmarks rather than model releases. Two are partial in opposite directions. FinMCP-Bench has code and an ICASSP 2026 paper but nothing on either model hub, while CARE has a paper and nothing else at all: the Code cell is a dash too. And M³FinMeeting is gated. It has no code, and its ModelScope and HuggingFace columns are merged into a single cell reading Application Required. ## LICENSES/ and NOTICE sit beside a root MIT Repository metadata records the licence as MIT, and a `LICENSE` file sits at the root. There is also a `LICENSES/` directory and a `NOTICE` file, both of which are standard companions when a repository aggregates work from several sources. For a hub that collects nine listed projects across seven directories, that combination means the per-project terms are not settled by the root licence. A `NOTICE` file exists to carry attribution, and a `LICENSES/` directory exists to carry more than one grant, and neither is needed if everything in the tree shares a single MIT text. The table itself does not have a licence column, so which terms apply to which sub-project is something you have to establish per directory rather than read off one page. For an academic or research audience this is a normal arrangement. For anyone planning to ship a sub-project inside a product, it is the first thing to check, and the answer is in those two entries rather than in the metadata. ## DianJin-SKILLS is 10 roles and 130+ skills across three financial lines The finance agent skill library was open-sourced on 2026.05.20 and named in the news as DianJin-SKILLS, with a Chinese name given alongside it. The scope is stated in one sentence: banking, insurance, and securities and asset management. The size is stated in the same sentence: 10 professional roles and more than 130 standardized skills, described as ready to plug into Agent frameworks. Two things are worth noting about how that is framed. It is a skill library rather than a model, so there are no weights to download, which is consistent with the table showing code only. And it is organised by job role first, ten of them, with skills underneath, so the unit of reuse is a workflow inside a role rather than a single capability. For an agent framework that reads skills from a directory, that shape matters more than the count. A library keyed to ten named roles is easier to map onto an existing agent's job description than a flat list of a hundred and thirty capabilities. ## DianJin-RED perturbs the user, the platform, or the tools The red-teaming benchmark was released on 2026.08.17 and is described as an action-grounded red-teaming benchmark for complete agent systems. The numbers are specific. There are 1,661 executable cases and 15 intervention strategies, and the document groups those strategies into three kinds: user input, agent-platform state, and external tools and data. That grouping is the design. Red-teaming an agent system rather than a model means testing three different seams. Perturbing the user probes how the system handles adversarial or mistaken requests. Perturbing agent-platform state probes what happens when memory, configuration or prior context is corrupted. Perturbing external tools and data probes whether the system trusts its own tools. The cases are executable and the service worlds are isolated, which means each case runs against a sandboxed stand-in for a real financial service rather than a recorded transcript. For a domain with real money behind it, that isolation is the difference between a benchmark you can run and one you can only read. ## Eight venue acceptances, and a news list instead of a changelog There is no CHANGELOG in the repository. The release history is the dated news list at the top of the README, and it is long enough to function as one. Counting the venues named across the entries gives eight: ACL 2024 for the CFLUE benchmark, ACL 2025 for M³FinMeeting, AAAI 2026 for the customer support work, ICASSP 2026 for both FinMCP-Bench and CARE, ACL 2026 main conference for CFMME, IJCAI 2026 for Fin-PRM, and Findings of EMNLP 2026 for FinGuard. Two small inconsistencies are visible in that list. The 2025.04.23 entry announces the DianJin-R1 series with two models, a 7B and a 13B, while the visible rows of the availability table show 32B and 7B. And CFMME appears in the news as an ACL 2026 acceptance but has no row in the part of the table shown here. The repository has no GitHub releases, and the last push was on 2026-08-28, three days after the FinGuard entry dated 2026.08.25. The repository is not archived. ## Conclusion Use qwen-dianjin if you are looking for financial-domain datasets, evaluation benchmarks or a finance skill library, and you are comfortable picking individual directories. Do not expect a single installable package, because the repository has no manifest and the primary language entry reflects the sub-projects rather than a root library. Before using anything here, check the availability row for your project, since several have code but no weights and one is application gated, and read the per-project terms under `LICENSES/` rather than relying on the MIT recorded at the root. ## FAQ ### What is Qwen DianJin? It is Alibaba Cloud's open source hub for financial AI research, described as LLMs for the financial industry. The repository collects released code, models and data across seven sub-project directories including DianJin-R1, DianJin-OCR-R1, Fin-PRM, DianJin-RED, DianJin-SKILLS, FinMCP-Bench and CSC, with no single installable package at the root. ### Which Qwen DianJin models and datasets can I download? The table points at both ModelScope and HuggingFace for Fin-PRM, DianJin-OCR-R1, the CSC datasets and the DianJin-R1 series, with 32B and 7B rows visible. DianJin-SKILLS and DianJin-RED are code only, CARE has a paper and nothing downloadable, and M³FinMeeting is marked application required rather than openly available. ### What is DianJin-RED? It is an action-grounded red-teaming benchmark for complete agent systems, released on 2026.08.17. It contains 1,661 executable cases and 15 intervention strategies grouped into user input, agent-platform state, and external tools or data, and it runs against isolated service worlds rather than recorded transcripts. ### What does DianJin-SKILLS contain? It is an AI Agent skill library for finance, open-sourced on 2026.05.20, covering banking, insurance, and securities and asset management. It is organised as 10 professional roles with more than 130 standardized skills, and is described as ready to plug into Agent frameworks. ## Sources - [aliyun/qwen-dianjin on GitHub](https://github.com/aliyun/qwen-dianjin) - [Issues](https://github.com/aliyun/qwen-dianjin/issues) - [License: MIT](https://github.com/aliyun/qwen-dianjin/blob/master/LICENSE) - [Project website](https://tongyi.aliyun.com/dianjin) - [README](https://github.com/aliyun/qwen-dianjin/blob/master/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/aliyun-qwen-dianjin