Weekly roundup
AI model drops: Pika H3 Flex and SFX v3
A busy end to the week for sound. Pika and ElevenLabs both shipped on Thursday, and both are worth an hour if you cut your own audio. SenseTime put out an odd but useful open model for motion graphics, and we've caught up with Tencent's Hy-Nomas from the week before. The language models, including Google's Gemini 4 Carbon and the open Mistral Large 4, are listed at the bottom.
Pika H3 Flex: picture with the sound already split
Pika took MiniMax's H3 and post-trained it to run in six steps instead of 50, so a 15-second clip with stereo sound at 768p can come back in seconds, from up to 12 image, video and audio references. The bit editors will like is in the API: you can ask for five separate stems (dialogue, score, SFX, Foley and ambience), so you can drop the music without losing the lines. Pika is releasing open weights too, but we haven't checked that licence yet, so read it before a paid job. The Model drops line has the link.
SFX v3: several sounds from one prompt
ElevenLabs' new sound effects model is better at prompts that describe more than one sound, so a crowd cheering, a ball dropping and the same crowd booing can come out of a single request. ElevenLabs also says it's cleaner than v2, with fewer artefacts on gongs, chimes and whistles, the sounds that usually give generated effects away. It's in ElevenCreative, including Studio and the Music Editor, and through the API. Worth trying for temp effects on a rough cut, keeping the ones that hold up.
SenseNova 6.8 Code2Video: motion graphics you can edit as code
This one doesn't make pixels: SenseTime's open 27B model writes the program that draws each frame, runs it, checks the result and fixes it, so you end up with code you can change rather than a clip you have to regenerate. SenseTime says it's for motion graphics, explainers and stylised 2D and 3D scenes, not photoreal footage, so title sequences and explainer inserts are the use to test. The model card lists Apache 2.0, but there's no licence file in the repo yet, so we've left the line untagged.
Hy-Nomas: steer the shot while it's still generating
Tencent released this on 21 Oct and it went up after last week's roundup, so here it is: an open-weights interactive video world model where you change the scene, the action or the camera with text as each new segment generates, at 720 x 1280 and 24 fps. An INT8 setup runs on one RTX 4090, and the code and weights are Apache 2.0, so commercial use is fine. It's a good one for previs, when you want to push a camera around a space without building a 3D scene.
Also dropped
- Mistral Large 4 Mistral AI, 29 Oct 2026
Open weights for Mistral's 1-trillion-parameter multimodal model (text and image in, text out), on Hugging Face and the Mistral API. - Gemini 4 Carbon Google, 28 Oct 2026
Google's new frontier Gemini model for coding, cybersecurity and long tasks; access starts with selected cyber defence partners. - LFM3 Liquid AI, 27 Oct 2026
Open-weight family of small 2B-class models on one backbone: text, vision, audio, omni and decision models that run on phones and laptops. - GPT-6.1 Terra OpenAI, 26 Oct 2026
Fast, cheaper GPT-6.1 model for visual reasoning, computer use and writing, with a 1.05M-token context; in ChatGPT Work, Codex and the API. - Kimi K3.1 Pro Moonshot AI, 24 Oct 2026
First model in the K3.1 series, with a 1M-token context and a faster Highspeed variant; live in Kimi and the Kimi API, full weights promised by 4 Nov. - Canva Craft Canva, 22 Oct 2026
Canva's first in-house image model: plans the layout first, then renders up to 2K across more than 30 aspect ratios. In Canva and the public API.
Related on AI Field Notes
- ElevenLabs: our model notes Model notes
- Film news, October 2026 News month
Every claim here comes from the maker's own page, linked above. Licence tags are what the maker's licence or terms said on the day we filed, not legal advice, so check the linked terms before client work. The full day-by-day list is on Model drops, and older roundups are in the weekly archive.