Lance, ByteDance’s newly open‑sourced 3‑billion‑parameter multimodal model, runs on a single 40 GB GPU, tops HuggingFace trend charts, and achieves leading scores on DPG Bench, GenEval, and video generation benchmarks while surpassing several state‑of‑the‑art single‑modal models.
Alibaba’s HappyHorse-1.0, a new text‑to‑video and image‑to‑video model from the ATH team, claimed the #1 spot on Artificial Analysis’s video arena rankings, matches ByteDance’s Dreamina Seedance 2.0 in Elo score, supports four generation modes, and will open its API around April 30.
MedGRPO, a multimodal large model, achieves a breakthrough in medical video understanding by introducing clinical semantic parsing that aligns visual cues with structured medical knowledge, boosting performance and raising ethical questions about AI’s evolving role from a supportive assistant to a collaborative clinical partner.
OmniLottie, a collaborative framework from Fudan, HKU, and Queensland University, uses a specialized tokenizer and a large multimodal model to compress Lottie files, generate vector animations from text, images or video, and achieves state‑of‑the‑art performance on custom benchmarks and extensive evaluations.
Alibaba's newly released Qwen3.5-Omni multimodal model adds an Audio‑Visual Vibe Coding feature that lets users describe a website or game with speech and gestures to generate code, while offering advanced audio comprehension, long‑duration media support, multilingual capabilities, fine‑grained voice control, and voice cloning, though its weights remain closed‑source.
Xiaomi unveiled MiMo‑V2‑Omni, a full‑modal agent base that unifies text, image, video and audio perception with tool‑calling and GUI actions, outperforming leading models such as Gemini 3 Pro and Claude Opus 4.6 on benchmarks, and offering a 256K‑context API for diverse real‑world tasks.
Qwen3.5‑397B‑A17B, the newly open‑sourced multimodal giant, combines a 400‑billion‑parameter sparse MoE architecture with FP8 pipelines and an asynchronous RL framework to deliver GPT‑5.2‑level capabilities, 60% lower memory usage, up to 19× higher throughput, and extensive image, video, and agent support, while outlining its deployment requirements and API pricing.
OPPO AI Center introduces AndesVL, an open‑source, fully‑adapted multimodal large model ranging from 0.6B to 4B parameters, designed for high‑performance, privacy‑preserving, low‑latency AI on mobile devices, with advanced architecture, training pipelines, on‑device optimizations, and state‑of‑the‑art benchmark results.
The open‑source UI‑UG‑7B multimodal model from Alipay combines UI understanding and generation in a single framework, delivering state‑of‑the‑art performance across referring, grounding, captioning, and code generation tasks while dramatically speeding up UI creation for developers.
IBM’s newly released open‑source model Granite‑Docling‑258M tackles the long‑standing challenge of converting diverse digital documents into machine‑readable, structured data by preserving layout, tables, formulas, and supporting multiple languages, while remaining lightweight at 258 M parameters and outperforming its predecessor SmolDocling‑256M‑Preview.
This article introduces FingER, a novel entity‑level evaluation framework and the FingER‑Instruct‑60k dataset for assessing AI‑generated video quality with fine‑grained reasoning, and demonstrates its state‑of‑the‑art performance on multiple benchmarks using advanced training strategies such as GRPO.
Cosmos Reason-1, a customizable multimodal physical AI model from NVIDIA, can be quickly deployed on Alibaba Cloud’s PAI‑Model Gallery with zero‑code, offering automatic cloud resource adaptation, ready‑to‑use APIs, enterprise‑grade security, and demonstrated superior reasoning on video tasks, while the upcoming tools enable fine‑tuning via SFT and RL.
Q-Insight, a multimodal large‑model introduced by Peking University and Volcano Engine, leverages reinforcement learning and a novel Group Relative Policy Optimization algorithm to evaluate image quality, providing detailed reasoning, degradation detection, and zero‑shot comparison, outperforming state‑of‑the‑art methods on multiple benchmarks.
DeepSeek‑VL2 is a state‑of‑the‑art multimodal model built on a Mixture‑of‑Experts architecture that combines a SigLIP‑L vision encoder with dynamic tiling, a two‑layer VL adaptor, and a DeepSeek‑MoE language model using Multi‑head Latent Attention, trained in three stages on diverse visual‑language and text data, and achieving strong results on benchmarks such as DocVQA and TextVQA, with full implementation and inference code available in PaddleMIX.
Alibaba's newly released Qwen2.5‑VL‑32B multimodal model delivers state‑of‑the‑art visual and textual performance, offering human‑aligned responses, superior mathematical reasoning, fine‑grained image understanding, and efficient deployment features that make it a compelling tool for developers and AI researchers alike.
Paris‑based startup Mistral AI has open‑sourced Mistral Small 3.1, a 2.4 billion‑parameter multimodal model that claims superior performance to OpenAI and Google equivalents, runs on modest hardware, processes up to 128 k tokens, and highlights a sustainable, accessible AI strategy.
UI‑TARS is a native GUI‑agent model that takes screenshots and natural‑language commands to predict the next UI action, and its integration with Midscene.js addresses the bottlenecks of generic multimodal LLMs, offering target‑driven planning, lower token usage, open‑source 7B/72B models, and detailed deployment guidance.
This article explores how compact multimodal models like OmniVision-968M enable efficient generative AI on edge devices, detailing their architectural advantages, benchmark superiority over larger models, and step‑by‑step instructions for local installation and visual inference using NexaSDK.
OPPO details how it deploys multimodal pretrained models on resource‑constrained edge devices by compressing CLIP‑based image‑text retrieval, adapting Chinese text‑to‑image generation with LoRA and adapters, and lightweighting diffusion models through layer pruning and progressive distillation, achieving sub‑3‑second generation while preserving cloud‑level quality.
