Moondream 1

Moondream 3 Preview is a compact frontier-oriented vision-language model built for fast visual reasoning, grounding, OCR, object detection, pointing, and structured output. It uses a 9B MoE architecture with 2B active parameters and extends context length to 32K, aiming to deliver strong real-world vision performance while staying efficient and inexpensive to run.

Moondream 2 is a small open vision-language model designed to run efficiently across many environments. It is the previous-generation Moondream model, released under Apache 2.0, and is positioned as a lightweight image-text model for practical multimodal use where smaller size and deployability matter more than maximum frontier scale.

Moondream 0.5B is a tiny open-source vision-language model built for edge devices and mobile platforms. With only 0.5B parameters, it is positioned as the world’s smallest VLM, designed for fast lightweight deployment on constrained hardware while still supporting practical real-world visual tasks.

MOSS-VL-Realtime is an 11B parameter vision language model built for continuous video streams. It perceives incoming frames and generates text in parallel, answers questions asked at any point in the stream, and can stay silent or revise earlier answers as new visual evidence arrives. It uses a cross attention architecture with timestamp aware frame encoding and a 256K context window.

Gemini 3 Pro is Google’s flagship Gemini 3 multimodal reasoning model for complex coding, agentic tasks, and long-context problem solving.

Qwen 3.6 Max Preview appears to be Alibaba’s newest preview flagship in the Qwen line, positioned above Qwen3.6-Plus and aimed at top-end coding, world knowledge, and instruction-following performance. I could not read the exact official blog page you linked directly, so I’m relying on current public reporting for this description.