Weekly roundup
This week in AI model drops: 20 Oct 2026
This is the first of these, so it looks back a bit further than a normal week. Most of what matters for filmmakers landed on Thursday 15 Oct, when Kling, ElevenLabs, Krea and Black Forest Labs all shipped on the same day. These are the four we'd actually open, and why.
Kling 4.0: up to 30 seconds in one take
Kling 4.0 went live for everyone on 15 Oct, and it'll now generate up to 30 seconds in a single pass, so a long take doesn't have to be stitched together from short clips. You can pin up to 10 keyframes along the way, and Omni Reference takes up to 15 items across images, videos, voice and elements, including white-model and depth-map videos to steer the camera. Kling lists 10-bit HDR at 1080p and 4K as coming soon, so don't plan a grade around it yet. Their own breakdown of The Beat walks through a 30-second single take shot by shot.
ElevenLabs Instant Soundtracks: a score and spot effects from your cut
Upload a video to ElevenCreative and it writes an original score, then watches the footage for moments that need a sound and puts effects on the timeline at those timestamps. The effects arrive as separate layers, so you can move or swap one without regenerating the music. You can also hand it your existing audio as a reference and it'll match things like tempo and instrumentation. On the API, separate_tracks gives you music and effects as separate files, for up to 10 minutes of video per request.
Krea 3: quick edits that keep the same face
Krea 3 is an image editing model, and it's quick: Krea says most edits come back in under 8 seconds at native 2K. It takes up to 6 reference images per edit, and their examples put the same cast into new poses, streets and light without them turning into different people. If you build stills for a character before taking frames into a video model, that's the part to test. It's $0.06 an image at 2K on krea.ai.
FLUX 3 Image Klein: open weights on your own card
Black Forest Labs put out FLUX 3 Image Klein as open weights: 5.6B parameters, text-to-image, editing and up to 10 reference images, at native 4K. BFL says the INT8 build runs on GeForce RTX cards with as little as 8GB of VRAM, through ComfyUI or Diffusers. Bounding boxes let you say where something goes in the frame, or keep an edit inside one area. The licence is free for commercial use under $5M annual revenue, so read it before you use it on a paid job.
Also dropped
- MiniMax M3.1 Flash MiniMax, 19 Oct 2026
New Flash language model for agents and coding; MiniMax says nearly three times faster online and 70% cheaper to use. - DeepSeek V4.1 Flash 1017 DeepSeek, 17 Oct 2026
V4.1 Flash updated to version 1017, now behind the deepseek-flash API name, with weights on Hugging Face. - OmniHuman 2.0 ByteDance (BytePlus), 15 Oct 2026
Audio-driven avatar video from one image plus speech, music or dialogue, now on the BytePlus API. - Muse Spark Open Meta, 15 Oct 2026
Open-weight multimodal model, 816B total and 42B active, 1M-token context, built for coding and agentic work. - Mirelo SFX 2.0 Mirelo, 14 Oct 2026
Video-to-sound model that replaces a clip's audio with synced effects, foley and ambience while keeping dialogue. - HeyGen Voice HeyGen, 14 Oct 2026
Voice cloning that turns a recording, a script and a target image into speech shaped for the scene on screen. - Drama 3 Fish Audio, 13 Oct 2026
Fish Audio's most emotive text-to-speech model, live on fish.audio, the Fish Audio app and the API. - Gemini 4 Argon Google, 13 Oct 2026
Google's frontier model now generally available in the Gemini API, AI Studio and the Gemini app for Pro and Ultra. - Wan 3.0 Prime Alibaba, 11 Oct 2026
Faster Wan 3.0 video model in Model Studio: text, image, video and audio in; 480P to 1080P, up to 30 seconds. - Claude Haiku 5.5 Anthropic, 7 Oct 2026
Anthropic's fastest small model yet, around 75% cheaper to run than Haiku 4.5 on average. - Vidu Q4 Preview Shengshu, 7 Oct 2026
Preview of the next Vidu model: image and reference-to-video with native audio, 3 to 16 second clips, 2K/4K out.
Every claim here comes from the maker's own page, linked above. The full day-by-day list is on Model drops, and older roundups are in the weekly archive.