Ant Ling 3.0 Flash VL

Ant Ling 3.0 Flash VL lists base rates of $0.021 per million input tokens and $0.062 per million output tokens. Estimate both input and output token volume and any additional charges for your workload. Novita BF16 via OpenRouter: $0.021/$0.0616 per MTok, cache read $0.0042, current 72% discount with no end date; 262,144 hosted context and 32,768 completion cap. DeepInfra FP16: $0.06/$0.18, cache $0.012, 131,072 context and same completion cap. Native Ant API is CNY: promotional ¥0.14/¥0.42, cache ¥0.028, versus regular ¥0.50/¥1.50, cache ¥0.10. Native checkpoint config is 131,072; the 256K recipe uses YaRN, not the default config. USD scalars identify the Novita route, not native CNY billing.

Ant Ling 3.0 Flash VL accepts about 131K tokens of context. That only matters if you routinely send very long documents, large codebases, or multi-turn histories that approach that limit.

Ant Ling 3.0 Flash VL is a budget tier from InclusionAI. It suits image and video understanding, visual coding workflows, evaluating multimodal agents. Thinking is enabled by default with temperature=0.6/top_p=0.95/top_k=20; no graded effort enum is established. Native Ant’s thinking toggle is documented only for text Flash, so checkpoint controls are not inferred for its VL API. First-party OpenAI accepts image/video; Anthropic accepts images, not video. Inputs must be base64: up to 40 JPEG/PNG images and a 32 MB request body; video is one MP4/MOV/WMV up to 30 seconds, sampled at 2 fps with 32-frame cap. The OpenRouter completion cap is not a native hard cap. Publisher Terminal-Bench settings use Terminus 2, 2-hour timeout and three-run mean; no unsupported comparable numeric scores are assigned. MIT is declared metadata; the attempted raw LICENSE file is missing.

Ant Ling 3.0 Flash VL publishes open weights, but self-hosting depends on the licence, hardware footprint, quantisation quality, and serving stack. A hosted API is often cheaper until you have measured throughput and concurrency on your own hardware.