Ling 3.0 Flash VL: a free, open vision model that reads video too
124B MoE, 5.5B active, 256K context, MIT weights on Hugging Face - plus Feyn's open promptable cutouts and Genspark's slide-only model at ~1/17 the cost.
Copy markdown
Free, MIT, and multimodal — including video
inclusionAI's Ling 3.0 Flash VL is a 124B-parameter MoE (just 5.5B active per token) that takes images and video across a 256K-token context, scoring 42 on Artificial Analysis's Intelligence Index — 4 points above the text-only Ling 3.0 Flash. Weights are MIT on Hugging Face, and NovitaAI serves it free on OpenRouter behind an OpenAI-compatible API, so you can add chart-reading, document parsing, and screen-driving to an app today at $0.
Genspark's Gen-1 Slides: a model that only makes decks
Genspark post-trained the open-weight MiniMax M3 with reinforcement learning inside its own slide-rendering loop to ship Gen-1 Slides, a model specialized for turning one prompt into a finished presentation. It runs about $0.30 per million input tokens versus Opus 5's $5 — roughly 1/17 the price, or ~$0.44 a deck — and it's now the default in Standard mode on every plan.
Feyn's MultiMatte: name what to keep, it cuts the rest
MultiMatte is an open, promptable background remover: type "the dog" and it alpha-mattes everything else away, keeping hair and translucent edges instead of a hard cutout. It's a 19.49M-parameter LoRA on Meta's SAM 3 (860M total) and scores 0.901 on the DIS-VD benchmark versus SAM 3's 0.667. Install with pip install nobg and load feyninc/multimatte — three lines to a production cutout.
Build this weekend: a $0 "what's on my screen?" agent
With Ling 3.0 Flash VL free on OpenRouter, pipe a screenshot or short screen recording into the vision endpoint and have it name the UI elements, read an error dialog, or summarize a chart — no per-call cost while you prototype. Because the API is OpenAI-compatible, you can swap to a paid host later without touching your code.