Ling 3.0 Flash VL goes live on OpenRouter

InclusionAI's new mixture-of-experts model debuted on OpenRouter yesterday, September 10, served by Novita Labs. Ling 3.0 Flash VL packs 124 billion total parameters with only 5.5 billion active, an architecture that keeps costs low while preserving a 262,000-token context window. Pricing is free for now.

The model adds native visual perception for image and video input alongside visual-agent capabilities that let it operate on graphical interfaces. It runs in two hybrid modes — instant for quick lookups and reasoning for multi-step tasks — and supports tool calling as a native part of inference.

OpenRouter reports a median throughput of 69 tokens per second and median latency of 2.38 seconds, both the best among providers running the model. Uptime sits at 99.98% and 99.93% respectively, and an automatic failover reroutes requests to a healthy provider when the primary errors out, provided the request filters allow it. Per-provider availability data is accessible via the Endpoints API.

Ling 3.0 Flash VL builds on the original Ling 3.0 Flash, the same 124B/5.5B MoE developed by InclusionAI, strengthening language capabilities while adding the visual perception layer. The release arrives without official benchmarks from the company, so practical evaluation will come from the community over the coming days.

A wide context window, zero cost, and dual run modes make the model a convenient candidate for rapid prototyping of multimodal applications, video analysis, UI automation, or combined vision-language tasks. Automatic failover and the availability-monitoring API simplify production deployment without building custom observability infrastructure.