Ling-3.0-flash-VL: MIT open image+video model, free on OpenRouter
Ant Group's InclusionAI ships a 124B MoE (5.5B active) that reasons over image and video — MIT weights, day-one vLLM/SGLang, and a free hosted endpoint.
Copy markdown
124B open MoE, only 5.5B active per token
InclusionAI's Ling-3.0-flash-VL is a 124B-parameter Mixture-of-Experts that fires just 5.5B params per token, over a hybrid KDA + Gated MLA backbone with a 256K context. A ViT encoder plus VideoRoPE route images and video through the same reasoning path, not a bolted-on caption step.
MIT weights, day-one vLLM and SGLang
The weights ship MIT-licensed in BF16 and FP8 now, with FP4/INT4 quants promised, and arrive with day-one serving in vLLM and SGLang. You can self-host a multimodal reasoner with no research-license gate and no wait for community ports.
Free to try on OpenRouter right now
A free hosted endpoint via Novita on OpenRouter gives you 262K input / 32K output tokens at roughly 78 tok/s, so you can prototype before committing GPUs. It's a time-limited free tier, so treat it as a sandbox, not production.
What it unlocks: an agent that acts on what it sees
It's tuned to read complex visuals, reason over that evidence, and drive interfaces — aimed at screenshot-and-click agents and video-understanding pipelines, not just alt-text. It scores 42 on the Artificial Analysis Intelligence Index v4.1.1, four points above the text-only Ling-3.0-flash.