# Vision-DeepResearch: a multimodal agent allowed dozens of search turns > Vision-DeepResearch is a research repository from Osilly that pairs a multimodal deep-research model, the VDR-Bench benchmark and a video agent. It publishes MIT-licensed training and evaluation code plus model weights, but it offers no install command, no tagged release and no serving path. **Osilly/Vision-DeepResearch** — [ICML 2026 & EMNLP 2026] Multimodal deep-research MLLM and benchmark. The first long-horizon multimodal deep-research MLLM, extending the number of reasoning turns to dozens and the number of search-engine interactions to hundreds. - Repository: https://github.com/Osilly/Vision-DeepResearch - Stars: 689 · Forks: 56 - Language: Python - License: MIT - Published: 2026-10-21 · Updated: 2026-10-21 · Language: en - Canonical page: https://hysenlabs.com/projects/osilly-vision-deepresearch ## Vision-DeepResearch trades one-shot answering for dozens of search turns Vision-DeepResearch is the official repository for a multimodal deep-research model, the VDR-Bench benchmark for visual and textual search, and a video extension of the same idea. The problem it targets is specific: an assistant that can see an image is normally allowed one attempt at answering, so a detail that is too small, too faint or simply out of frame ends the exchange. The stated aim is to push the number of reasoning turns into the dozens and the number of search-engine interactions into the hundreds, so the model can keep pulling context after it has already started answering. That reframes the image from a static input into the starting point of an investigation. The people who should care are researchers benchmarking multimodal agents and engineers deciding whether a search-capable vision model belongs in their own stack. The code is Python and the licence is MIT, so the training and evaluation scripts can be read and reused without a commercial negotiation. ## One repository, three artefacts: model, benchmark, video branch This is not a single package. The top level carries a LICENSE file, a README and separate directories for the training stacks and the assets: Megatron-LM/, mbridge/, ms-swift/, rllm/, data/, checkpoints/, figs/ and Video-DeepResearch/, plus a .github/ folder. Three papers are attached to that structure, one for the model, one for the benchmark and one for the video agent, whose code the timeline says lives under Video-DeepResearch/. The evaluation assets sit in a nested spot, rllm/eval/README.md, and the timeline entry from 2026-05-24 points readers there for the released benchmark subset and detailed evaluation guidance. A reader who expects a pip-installable library with one entry point will not find one. What the directory shape does give is an accurate picture of the effort involved, and the acceptance by ICML 2026 recorded on 2026-05-01 is worth knowing when judging how seriously the surrounding work is being taken. ## The Direct Answer row shows where closed models still stand The performance table opens with a block labelled Direct Answer, and it lists GPT-5 at an average of 37.4, Gemini-2.5 Pro at 37.7 and Gemini-2.5 Flash at 30.1, then Claude-4-Sonnet at 21.3, Claude-3.7-Sonnet at 22.8 and the Qwen3-VL family from 19.2 for Qwen3-VL-8B-Instruct up to 24.1 for Qwen3-VL-30B-A3B-Thinking. Individual columns are harsher on some models than others: Claude-4-Sonnet records 2.0 on VDR and 4.0 on both MMSearch+ and MMSearch, while Gemini-2.5 Pro scores 8.0 on VDR but 39.8 on MMSearch. Read the numbers as one setup rather than a ranking of the models themselves. The block that would name the project's own agents is cut off, and no search backend, turn budget or judge model is described, so the table cannot be reproduced from the text alone. That is a documentation gap, not evidence that the results are wrong. ## Getting the artefacts means Hugging Face, since no install step exists There is no installation command to copy. The project distributes its artefacts on Hugging Face and ModelScope instead: Vision-DeepResearch-30B-A3B, described as SFT plus RL, and Vision-DeepResearch-8B, described as SFT-only, alongside the dataset pages Vision-DeepResearch-Eval, Vision-DeepResearch-Toy-SFT-Data, Vision-DeepResearch-Toy-RL-Data, VDR-Bench and VDR-Bench-testmini. The video agent appears on ModelScope as Video-DeepResearch-30B and Video-DeepResearch-35B-A3B. The three papers sit on arXiv, and a project page is linked from the repository header. One line in the documentation still says the datasets, code and weights will be released, while the timeline lists the 30B-A3B model on 2026-06-08 and the toy datasets plus the 8B model on 2026-02-02, so treat the release state of any given artefact as something to confirm on the host site rather than something the overview guarantees. For a first real use, the honest path is to pull a checkpoint, follow rllm/eval/README.md for the evaluation setup, and read the papers for the method before assuming a running pipeline is one command away. ## Training code splits across ms-swift and rllm directories The training code is split across two ecosystems, and the split tells you something. Supervised fine-tuning scripts live under ms-swift/run, which points at the ms-swift training framework, while reinforcement learning code lives under rllm/vision_deepresearch_async_workflow, a directory name that describes an asynchronous workflow. Megatron-LM/ and mbridge/ sit at the top level as well, the usual building blocks for large model training, and a checkpoints/ directory shows where training output is expected to land. The timeline records the SFT and RL code release on 2026-02-03 with a note that guidance would follow later, and a later entry on 2026-05-24 releases the benchmark subset and the detailed evaluation guidance. The practical consequence is that reproducing the models means assembling three or more frameworks, and the documentation does not state versions, dependency pins or hardware configuration for any of them. ## A RAG pipeline is the real alternative, and it wins some columns The table also contains a RAG Workflow block, and it is the closest thing to a head-to-head comparison available. On MMSearch it records 43.9 for Gemini-2.5-flash, 32.7 for Claude-3.7-Sonnet and 29.2 for Qwen-2.5-VL-72B, and on LiveVQA 41.3, 30.3 and 35.7 respectively, with BC-VL far lower at 12.1, 10.0 and 10.2. The difference in approach is what matters. A conventional retrieval pipeline pulls text and answers once, so its quality is bounded by the chunk it retrieved. The deep-research setup lets a multimodal model plan, look again and search repeatedly, which is how it can chase a visual detail that no single chunk would carry. The weak BC-VL scores for the RAG row point at the reverse case: for questions one retrieval answers well, a long-horizon loop adds cost and can add error, so the plain pipeline is the correct tool more often than the framing around this project suggests. ## No GitHub releases and no serving instructions leave real gaps Two facts define the upgrade path, and both are awkward for production use. First, the repository has no GitHub releases, so there is no tagged version to pin and no changelog to read; the timeline is the only version history on offer. The last push was on 2026-08-08 and the repository is not archived, so it has not been abandoned, but that is the strongest maintenance claim the evidence supports. Second, there is no documented serving path. The repository exposes no HTTP API, no command line entry point and no configuration schema for pointing the agent at a particular search engine, and the top level has no requirements file, so a deployment team assembles the environment itself. The MIT licence removes the usual friction around reuse, and it says nothing about the obligations attached to the underlying weights or to the data behind them. ## Conclusion Researchers comparing long-horizon multimodal agents and teams who need a vision model that keeps searching will find this repository worth reading, and the MIT licence plus published weights make an evaluation cheap to start. Teams that need a stable HTTP API, a pinned release or a documented search backend will not find those here, and the missing install and serving instructions turn production adoption into a research task rather than an integration. Before deciding, check three things on the artefact you intend to use: the release state of the checkpoint on Hugging Face, the evaluation guidance in rllm/eval/README.md, and the dependencies behind ms-swift/ and rllm/ that nobody has pinned yet. The last push was on 2026-08-08 and the repository publishes no GitHub releases, so pin the exact commit you evaluate. ## FAQ ### How do I install Vision-DeepResearch? There is no install command in the repository. The project points to Hugging Face for the Vision-DeepResearch-30B-A3B and Vision-DeepResearch-8B checkpoints and the related datasets, to ModelScope for the video agent, and to rllm/eval/README.md for the evaluation setup. ### What is VDR-Bench and how do I run the evaluation? VDR-Bench is the benchmark published alongside the model as VDR-Bench and VDR-Bench-testmini, with a smaller released subset named Vision-DeepResearch-Eval. The detailed evaluation guidance lives in rllm/eval/README.md, and the documentation does not state the versions or configuration needed to reproduce the table of scores. ### Is Vision-DeepResearch still being worked on? The repository is not archived and the last push was on 2026-08-08, so the work is recent. It publishes no GitHub releases, so there is no tagged version to pin and the timeline is the only version history. ### Can Vision-DeepResearch handle video? That is the separate Video-DeepResearch line of work, which extends the same idea to continuous video streams. Its paper was released on 2026-08-05, its code lives under Video-DeepResearch/, and the model appears on ModelScope as Video-DeepResearch-30B and Video-DeepResearch-35B-A3B. ## Sources - [Issues](https://github.com/Osilly/Vision-DeepResearch/issues) - [License: MIT](https://github.com/Osilly/Vision-DeepResearch/blob/main/LICENSE) - [Osilly/Vision-DeepResearch on GitHub](https://github.com/Osilly/Vision-DeepResearch) - [README](https://github.com/Osilly/Vision-DeepResearch/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/osilly-vision-deepresearch