Clinical laboratory microscopy vision-language model that identifies cells, describes their morphology and drafts concise reports with box locations.
No providers recorded yet. Browse all providers
Reading a stained smear is not a single classification. A laboratory scientist finds the cell worth attention, names it, explains which nuclear and cytoplasmic features justify the name, says where it sits, and then answers a colleague's follow-up question in the house terminology and report format. A model can get the morphology right and still be hard to use if it ignores the requested format, rambles when a one-line report was asked for, or loses the thread on the second question. Lingjian-1.5 targets that gap. It is a successor release to Lingjian, the laboratory-morphology vision-language model whose Lingjian-1.0-Chat checkpoint was described and evaluated in npj Digital Medicine. Version 1.5 keeps the same scope and output style and is trained to follow instructions and hold a dialogue more reliably.
Lingjian-1.5 was released in June 2026 by the LingJian AI Research Team, from the same Hugging Face account that published Lingjian-1.0-Chat. That earlier checkpoint came from a group led at Jiulongpo District People's Hospital in Chongqing. The 1.5 model card names no institution of its own and gives no author list.
The release has no paper of its own; its card lists a technical report as in preparation. The improvements the card claims over 1.0-Chat are qualitative: better recognition of clinically relevant findings with fewer omissions in descriptions, closer adherence to requested style, detail level and terminology, and more natural multi-turn conversation. No benchmark results for 1.5 have been published. The accuracy and reader-study figures in the npj Digital Medicine paper were measured on Lingjian-1.0-Chat, not on this checkpoint.
/bbox to an instruction asks the model to mark key targets, and it writes the box coordinates into its reply instead of producing them from a separate detection head.Lingjian-1.5 uses the Qwen3-VL architecture and is fine-tuned from Qwen3-VL-8B-Instruct by way of internal Lingjian intermediate checkpoints. The released configuration matches the 8B Qwen3-VL layout: a 36-layer language model with 4,096-dimensional hidden states and grouped-query attention (32 query heads, 8 key-value heads), and a 27-layer vision encoder with 16-pixel patches whose features from layers 8, 16 and 24 are injected into the language model. The safetensors weights hold about 8.77 billion BF16 parameters.
The published training arguments show one epoch of full-parameter fine-tuning with MS-Swift. Training started from an internal stage-1 Qwen3-VL-8B checkpoint, with a learning rate of 1e-5 for the language model and 1e-6 for the vision tower, and a maximum sequence length of 4,000 tokens. The data mixture has 21 files: long-form image captions, grounding and local bounding-box sets (several of them augmented), visual question answering, reasoning-style QA, laboratory text QA, self-cognition data and a 100,000-example general chat set. Validation used a held-out set of 462 concise-answer items. None of these datasets are released, and the card gives no image counts or sources for them.
The card frames Lingjian-1.5 as an auxiliary tool for clinical laboratory staff. Its intended uses include second review of microscopy images, cell identification and morphology description, flagging abnormal findings and drafting short reports, laboratory quality control, and morphology teaching. It also lists research uses: screening datasets, assisting annotation and expert review, fine-tuning for a specific laboratory subdomain, and integration with laboratory information systems. The card excludes primary diagnosis without qualified supervision, fully automated or emergency clinical decisions, and pixel-level detection, where dedicated detection models are more precise. It also asks every deployment to validate the model locally across sample types, staining protocols and instruments.
Lingjian-1.5 shows a clinical-laboratory model family being maintained after its paper came out. The predecessor was evaluated externally, and this release concentrates on usability: format control, terminology and dialogue. The weights and training configuration are openly released under Apache-2.0, which makes the checkpoint easy to test or adapt. Its standing rests on the predecessor's evaluation until the promised technical report appears, though. Laboratories adopting it need their own validation, and the card itself warns that the model may produce plausible but incorrect statements on ambiguous images, rare cell types or unusual staining.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.