vision-language

Microsoft has published Mage-VL-Cowork on Hugging Face, a 4B-parameter vision-language policy that combines GUI control and code/terminal execution in one checkpoint. According to the model card, it reaches 66.0% on OSWorld-Verified under a 50-step budget, the best among 4B models. It is released under Apache-2.0.

Liquid AI has published d1-2B, a 1.92B-parameter decision model built on LFM3-VL-2B with a 128K-token context. It returns typed, calibrated answers in a single forward pass without generating tokens, and Liquid AI claims 5.7 ms per decision on an NVIDIA H100.

Microsoft has released Mage-VL-Cowork, a 4B-parameter vision-language policy that combines GUI control and terminal execution in one model. Microsoft claims it is the best 4B model on OSWorld-Verified, at 66.0% under a 50-step budget. Weights are published under Apache-2.0.

A research team involving miHoYo's Honkai AI researchers and university collaborators has open-sourced GameScaling, a family of eight checkpoints from 800 million to 27 billion parameters. The models generate keyboard and mouse actions from visual observations and were built by continued pretraining of Qwen3.5 vision-language models.

Tencent's WeChat Vision team released WeVisOne-VL-4B, a vision-language model built on Qwen3-VL that handles object-level and pixel-level perception. It adds a 9.6M-parameter pixel decoder to produce segmentation masks and full-resolution alpha mattes.

Tencent's WeChat Vision team released WeVisOne-VL-2B, a 2B-parameter vision-language model built on Qwen3-VL. It handles object-level and pixel-level perception, including referring segmentation and alpha matting, through structured next-token prediction and a 9.5M-parameter pixel decoder.

OrcaRouter has published OrcaCyber Nano 1.0, a 27B vision-language reasoning model for defensive security, as GGUF files for local inference with llama.cpp. It supports a 262,144-token context window and text, image and video input. Benchmark scores are self-reported, and the harness details are not yet published.

NVIDIA has published Nemotron 3.5 Super VL in NVFP4 quantization on Hugging Face: a 120B-parameter, 12B-active hybrid Mamba-2/Transformer MoE with a 1M-token context window and text, image and video input. NVIDIA reports the quantized checkpoint stays close to BF16 on most benchmarks. It fits on one DGX Spark, one H200, two H100s or one GB200.

OpenBMB has released MiniCPM-V 4.7 35B-A3B, the first mixture-of-experts model in the MiniCPM-V series, with open weights on Hugging Face. The company claims it runs at 13+ tokens/s using 3 GB of memory on an iPhone 17 Pro Max, and that it beats Qwen3.6 35B-A3B on overall score in less inference time in Instruct mode.

StepFun has released Step-3.7-Flash, a 198B-parameter sparse Mixture-of-Experts vision-language model that activates approximately 11B parameters per token. The model supports a 256K context window, native image understanding via a 1.8B-parameter vision encoder, and offers three selectable reasoning levels.

Perceptron has released Perceptron Mk1, a vision-language model designed for video understanding and embodied reasoning tasks. The model accepts image and video inputs with 33K context window, priced at $0.15 per 1M input tokens and $1.50 per 1M output tokens, and supports structured spatial annotations on demand.

NVIDIA released Nemotron 3 Nano Omni, a 30B-A3B Mixture-of-Experts model that processes text, images, video, and audio. The model uses a hybrid Mamba-Transformer architecture with 128 experts and achieves 65.8 on OCRBenchV2-En and 72.2 on Video-MME, while delivering up to 9x higher throughput on multimodal tasks compared to alternatives.

NVIDIA engineer Asier Arranz demonstrated Gemma 4 running as a vision-language agent (VLA) on a Jetson Orin Nano Super with 8GB RAM. The model autonomously decides when to access a webcam based on user queries, with no hardcoded triggers—performing speech-to-text, vision analysis, and text-to-speech entirely locally.

Researchers introduced ProactiveBench, a benchmark testing whether multimodal language models ask for help when visual information is missing. Out of 22 models tested—including GPT-4.1, GPT-5.2, and o4-mini—almost none proactively request clarification, instead hallucinating or refusing to respond. A reinforcement learning approach showed models can be trained to ask for help, improving performance from 17.5% to 37-38%, though significant gaps remain.

Zhipu AI released GLM-5V-Turbo, a multimodal coding model that converts design mockups directly into executable front-end code. The model processes images, video, and text with a 200,000-token context window and 128,000-token max output, priced at $1.20 per million input tokens and $4 per million output tokens.

TIIUAE has released Falcon Perception, a 0.6B-parameter early-fusion Transformer that combines image patches and text in a single sequence for open-vocabulary object grounding and segmentation. The model achieves 68.0 Macro-F1 on SA-Co (vs. 62.3 for SAM 3) and introduces PBench, a diagnostic benchmark that isolates performance across five capability levels. TIIUAE also released Falcon OCR, a 0.3B model reaching 80.3 on olmOCR and 88.6 on OmniDocBench.

The Allen Institute for AI has released MolmoWeb, a fully open web agent that operates websites using only screenshots without access to source code. The 8B-parameter model achieves 78.2% success on WebVoyager—nearly matching OpenAI's o3 at 79.3%—while being trained on one of the largest public web task datasets ever released.

Reka has released Reka Edge, a 7-billion parameter multimodal model designed for efficient image and video understanding. The model features a 16,384 token context window and is priced at $0.20 per million input and output tokens.

Alibaba's Qwen team has released Qwen3.5-4B, a 4 billion parameter multimodal model capable of processing both images and text. The model is available on Hugging Face under an Apache 2.0 license, making it freely available for commercial and research use.