10 Open-Weight Multimodal Models to Know in 2026

Updated August 27, 2026

TL;DR: “Multimodal” now covers more than image chat. Current open-weight models can read documents, reason over charts, follow visual interfaces, analyze video, process audio and, in some cases, speak while watching a live stream. There is no universal winner: the right checkpoint depends on modality, license, context length, hardware, latency and performance on your own data.

Many models are called open source simply because their weights can be downloaded. That is too broad. A fully open model also provides enough training data, code and methodology to study and reproduce the system. An open-weight model exposes its trained weights but may use a custom license or keep parts of its training pipeline private.

This guide therefore uses open-weight as the umbrella term. Several entries – including Qwen3.8-27B, Gemma 4, GLM-5.3-Flash, MiniCPM-o 4.5, Molmo 2, STEP3-VL-10B and Muse Glimmer – use permissive Apache 2.0 or MIT licenses. Kimi K3 uses its own model license, and InternVL checkpoints can inherit terms from their underlying components. Always check the exact model card and acceptable-use policy before commercial deployment.

Qwen3.8-27B is Alibaba’s current deployment-oriented dense multimodal model. It has 27B parameters and natively handles text, images and video, with capabilities spanning documents, charts, visual coding, GUI interaction and long-horizon agentic tasks.

The released weights support a native 262,144-token context window and can be extended to 1 million tokens with the documented YaRN configuration. The checkpoint is Apache 2.0. Qwen3-VL remains widely deployed, but it is no longer Alibaba’s current vision-language generation.

Best fit: teams that want one current, general-purpose model for visual reasoning, documents, computer use and video without moving to a trillion-parameter system.

Gemma 4 spans small effective-parameter models for edge and browser deployment, a 31B dense model, a 26B-A4B mixture-of-experts model and the 12B Unified model. The family supports image understanding, OCR, documents, charts, video frames, pointing, function calling and long context. Audio input is available on the E2B, E4B and 12B Unified variants.

Context reaches 128K for the smaller variants and 256K for the medium and larger releases. Google publishes Gemma 4 under the Apache 2.0 license and also provides a separate prohibited-use policy.

Best fit: teams that need one family across mobile, workstation and small-server deployments. E2B and E4B provide compact audio-capable options; use 12B Unified when you need a larger shared text, image and audio model.

Kimi K3 is Moonshot AI’s current open-weight native multimodal agentic model. It uses a 2.8-trillion-parameter mixture-of-experts architecture with 104B activated parameters, a 401M-parameter MoonViT-V2 encoder and a 1-million-token context window.

The model is designed for long-horizon coding, research, tool use and visually grounded work, and Moonshot documents native understanding of text, images and video. The weights use the custom Kimi K3 License rather than the Modified MIT terms used for Kimi K2.5 and K2.6.

Best fit: frontier-scale multimodal agents running on substantial infrastructure. The sparse architecture reduces active compute, but this is still not a practical laptop or single-workstation model.

GLM-5.3-Flash is Z.ai’s first natively multimodal model in the GLM-5 series. Released on August 26, 2026, it has 320B total parameters and activates about 18B per token, combining sparse and linear attention with native visual understanding.

The model targets coding, agents, browser and computer use, and visually grounded knowledge work. Z.ai publishes the weights under the MIT license and documents local serving through Transformers, vLLM, SGLang, TokenSpeed and KTransformers. GLM-4.6V remains a useful smaller-family reference, but it is no longer the current GLM multimodal release.

Best fit: large multimodal agent and coding workloads that need strong visual reasoning with lower active compute than the model’s total parameter count suggests.

MiniCPM-o 4.5 is a 9B end-to-end omni-modal model that accepts text, images, video and audio and produces text and speech. Its defining feature is full-duplex streaming: it can continue receiving live video and audio while generating speech and text.

OpenBMB also provides quantized formats and support for Transformers, vLLM, SGLang, llama.cpp and Ollama. Both code and weights are Apache 2.0.

Best fit: private, on-device assistants that need to see, listen and speak in real time. Evaluate latency on the target device; “on-device” does not mean every 9B configuration will fit every phone or laptop.

Molmo 2 is Ai2’s current multimodal family for image, multi-image and video understanding. It comes in 4B and 8B variants plus Molmo 2-O 7B, which pairs the vision system with Ai2’s fully open OLMo language backbone.

Its distinctive strengths are grounding, pointing, counting and tracking across images and video. Ai2 publishes weights, code, training recipes and new datasets; the checkpoints use Apache 2.0.

Best fit: research and products that need spatially grounded answers, video tracking or a more transparent training stack. Molmo 2-O is the clearest choice here when openness itself is a requirement.

InternVL3.5 offers one of the widest checkpoint ranges: dense models from roughly 1B to 38B and sparse variants up to 241B-A28B. It targets general image understanding, documents, video and multimodal reasoning, with both native project formats and Hugging Face-compatible releases.

The main project is MIT-licensed, but some checkpoints incorporate language backbones or components with their own terms. License review must therefore happen at the checkpoint level rather than only at the repository level.

Best fit: teams that want to benchmark the same multimodal recipe across many compute budgets or fine-tune a size closely matched to their hardware.

STEP3-VL-10B is a 10B dense model released in 2026. It focuses on visual perception, STEM reasoning, OCR, documents, spatial understanding and GUI grounding, and it uses parallel coordinated reasoning as an optional test-time strategy.

The model and code use Apache 2.0. The official implementation currently emphasizes BF16 inference, so its parameter count should not be mistaken for minimal memory use without quantization support.

Best fit: teams looking for a compact, reasoning-oriented VLM for screenshots, interfaces, charts and technical imagery.

Muse Glimmer is Meta Superintelligence Lab’s 30B dense multimodal agentic model for local deployment. It combines a roughly 29.6B-parameter language model with a dedicated 1.8B-parameter perception encoder and supports text and image input with text output.

The model has a 131,072-token context window and is released under Apache 2.0, with official BF16, GGUF and ExecuTorch builds. Audio is not supported, and video is processed as individual frames rather than through a dedicated temporal architecture.

Best fit: local document analysis, coding and tool-using assistants where privacy and deployability matter. It is more practical than frontier-scale MoE models, but it is not a drop-in replacement for an existing Llama 4 stack.

PaddleOCR-VL-1.6 is a document-parsing pipeline, not simply a standalone 0.9B model. Its full page-level workflow pairs PP-DocLayoutV3 for layout analysis with the PaddleOCR-VL-1.6-0.9B recognition model, then handles region cropping, reading order and result assembly.

The pipeline extracts text and structure from PDFs and images and supports complex layouts, tables, formulas, charts, seals and multilingual documents. The PaddleOCR stack is Apache 2.0 and supports local and server deployment through multiple inference backends.

Best fit: OCR and structured document ingestion. Use the complete pipeline for page-level parsing; calling only the 0.9B recognition checkpoint does not reproduce the full system’s layout and reading-order capabilities.

Do not rank these models from vendor benchmark tables alone. Test the released weights with the same image resolution, frame sampling, prompt format, quantization and inference engine you plan to use in production. A small private evaluation set is more useful than a one-point lead on a saturated public benchmark.

A multimodal large language model processes language together with other data types such as images, video or audio. Some models only accept multimodal input and generate text; omni-modal systems can also generate speech or other modalities.

Open-weight models publish trained weights, but they may use custom licenses or withhold training data and recipes. Fully open-source AI provides the code, data and methodology needed to study and reproduce the system under open terms.

There is no universal winner. Qwen3.8-27B is a versatile general-purpose option; Kimi K3 and GLM-5.3-Flash target frontier-scale agents; Gemma 4 spans compact hardware tiers; MiniCPM-o 4.5 adds live speech; Molmo 2 specializes in grounding and tracking; Muse Glimmer targets local agents; and PaddleOCR-VL specializes in document parsing.

Yes. Small and quantized checkpoints can run on workstations and some laptops or mobile devices. Feasibility depends on precision, context length, image resolution, number of video frames, available RAM or VRAM, and support in the chosen inference engine.

Build a private test set that reflects your images, documents, languages and failure costs. Measure accuracy, hallucination, grounding, latency, memory use and robustness while keeping the prompt, resolution, frame sampling, quantization and inference engine consistent.