Best Phrase Grounding Models

Mistral Large 4, nicknamed Le Chonk, is a natively multimodal mixture-of-experts model from Mistral that accepts interleaved text and image input and produces text output. It uses a granular MoE design with roughly 1.05 trillion total parameters and 49 billion active per token, reported as 52 billion when embeddings and output layers are counted, paired with a 1.6 billion parameter vision encoder and a context window of one million tokens. The model is trained from scratch on about 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centers and supports more than 160 languages. It behaves as a hybrid instruct and reasoning system, with a reasoning effort setting that selects between direct answers and longer deliberation, alongside function calling and structured output for agentic workflows.Image understanding is a focus of this generation, covering documents, charts, technical drawings and natural scenes, and the model emits bounding box coordinates for visual grounding queries. Reported grounding results include 42 percent on Dense200 and 73 percent on the DIOR-RSVG remote sensing benchmark. Mistral describes agentic vision workflows in which the model zooms into gigapixel satellite imagery or engineering drawings to verify details, and reports coding results such as 62 percent on DeepSWE. Figures published at preview time are preliminary because the reinforcement learning phase is still in progress.

Qwen3.6-35B-A3B is a sparse Mixture-of-Experts (MoE) multimodal language model developed by the Qwen team at Alibaba Group. It carries 35 billion total parameters but activates only approximately 3 billion per forward pass via a learned routing mechanism, giving it the representational capacity of a large dense model at a fraction of the inference compute. The model is natively multimodal, processing images, documents, and video alongside text as a core architectural capability rather than an add-on. It supports a native context window of 262,144 tokens, extensible up to 1,010,000 tokens via YaRN. A key design feature is the unified thinking/non-thinking mode framework: users can switch between deliberate chain-of-thought reasoning and fast direct responses within a single model, and a "thinking preservation" option retains reasoning context across multi-turn agentic workflows to reduce redundant computation.The model is specifically optimized for agentic coding tasks, including repository-level reasoning, frontend workflow generation, multi-step tool use, and MCP (Model Context Protocol) integration. On SWE-bench Verified it scores 73.4%, on Terminal-Bench 2.0 it scores 51.5%, and on MCPMark it scores 37.0%. For vision-language tasks it achieves 92.0 on RefCOCO, 89.9 on OmniDocBench 1.5, and 83.7 on VideoMMMU. The model also supports Multi-Token Prediction (MTP) for speculative decoding. All Qwen3.6 open-weight models are released under the Apache 2.0 license.

Florence-2, introduced by Microsoft Research at CVPR 2024, is an open-source vision-language foundation model designed to unify diverse computer vision tasks within a single sequence-to-sequence framework. Unlike traditional models that specialize in specific tasks, Florence-2 accepts both images and text prompts and outputs text for tasks such as captioning, object detection, segmentation, OCR, and region-based grounding. It comes in two sizes—Florence-2-base (~230M parameters) and Florence-2-large (~770M parameters)—and is trained on FLD-5B, a large dataset of ~126M images with ~5.4B annotations.The model demonstrates strong zero-shot and fine-tuned performance, often rivaling larger vision-language systems while remaining lightweight and efficient. Released under the MIT license, all weights are publicly available, making it accessible for fine-tuning and deployment in applications like VQA, content tagging, accessibility, and research. Florence-2’s compact design, versatility, and openness position it as a practical alternative to larger proprietary multimodal models.

YOLO-World v2 Small (YOLO-World-S-v2) is the smallest variant of Tencent AI Lab’s YOLO-World v2 family, released around February 2024 under GPL-v3. With ~13 million parameters, it adopts a prompt-then-detect paradigm using offline vocabularies and is pretrained on large-scale datasets such as Objects365 and GoldG. The model processes image inputs at 640×640 or 1280×1280 resolutions and supports zero-shot open-vocabulary object detection, enabling recognition of novel categories from text prompts without retraining.Evaluations show competitive results across benchmarks like LVIS and COCO, while maintaining real-time efficiency. On an NVIDIA V100, the small variant reaches ~74 FPS at standard resolutions. Together with larger YOLO-World v2 models, it provides a scalable framework for efficient, open-vocabulary detection across diverse deployment settings.