VLX-Seek is a fine-grained perception vision-language model from OM AI Lab (om-ai-lab), built for edge-side and embodied scenarios where a system must not only describe an image but locate specific objects, tell instances apart, and recognize when a requested target is simply absent. Its tagline, "From Coordinate Generation to Region Reference," names the architectural bet: rather than having the language model emit raw numeric coordinates like [x1, y1, x2, y2] for every object, VLX-Seek reformulates localization as retrieval over a set of candidate regions. A region proposal stage first recalls candidate foreground boxes using a proposal network decoupled from the VLM backbone, so it can be swapped for another detector, user-supplied boxes, or visual prompts. Each candidate is then turned into a region-level representation by a Hybrid Fine-Grained Region Encoder (HFRE), which combines a semantic pathway that preserves the base VLM's visual-language alignment with a detail pathway carrying higher-resolution local structure, boundaries, and small-object cues, plus a SimpleFP module for multi-scale ViT-style features. Those features are projected into the LLM's embedding space so each box becomes a token the model can address directly, serialized as special tokens like <obj0>, <obj1>, <obj2>. Asked to find "people wearing red," the model answers <ground>people wearing red</ground><objects><obj2><obj5></objects> instead of writing coordinates from scratch, which the README argues is more compact, easier to parse, and faster to decode when several objects are involved. The same mechanism covers open-vocabulary detection, referring expression comprehension, region OCR, region captioning, counting, visual region reasoning, and general VQA that needs no proposals. Training runs in two stages: a region-language alignment stage that trains HFRE, the region-language connector, and the new special tokens while keeping most of the base VLM frozen, then perception instruction tuning that mixes in general VLM data to limit catastrophic forgetting and adds explicit negative/rejection samples so the model reports "no match" rather than hallucinating a box. Two disclosures matter for reading the results. The README states that, due to company policy, the internally trained object proposal network used in the team's blog post is not released; the open repository substitutes the open-source lightweight WeDetect-Base-Uni detector as the proposal source instead. And the benchmark charts shipped in the README (MSCOCO, ODinW13, RefCOCO/RefCOCO+/RefCOCOg, PixMo Count) are figures rather than a numeric table, and the README says they report the original VLX-Seek 3B model, pointing readers to the team's blog for 1.5-series numbers. VLX-Seek 1.5 was announced in a July 2026 blog post and the inference code plus the 10B checkpoint were open-sourced on 2026-07-23; the 1.5 line is planned as a 0.6B/3B/10B family, but only VLX-Seek-1.5-10B (Hugging Face omlab/VLX-Seek-1.5-10B, Apache-2.0) has shipped weights so far. The project builds on OM AI Lab's earlier open releases OmDet-Turbo, VLM-R1, and VLM-FO1 as the latest entry in what the team calls its VLX series.