Blog ·

New open-source AI models this week (November 2026) and where to try them

Nanbeige4.3-33B, Apex 35B, Groot 19B, Unlimited OCR 2, Paradee, Audio8 and more from this week's AI news: what each one does and a link to run it online.

This week's AI news brought a long list of new open models. Here are the ones you can already run in your browser on LiveWeights, grouped by what they do, with license and size. Each one has a free first try.

Chat and reasoning

Nanbeige4.3-33B

A 33B open model for reasoning and chat, competitive with much larger models in its size class. Turn reasoning on for maths, logic and multi-step questions. Apache-2.0.

Nanbeige4.3-33BNanbeige · Apache-2.0A 33B open model for reasoning and chat, competitive with much larger models at its size class.Try it online →

Apex 35B

A 35B mixture-of-experts model that decides on its own when to think and for how long; you can open its reasoning under each answer. Apache-2.0.

Apex 35BApex Intelligence · Apache-2.0A 35B mixture-of-experts model that decides on its own when to think and for how long.Try it online →

LFM3 1.7B and Spark-X2.6 1.7B

Two tiny models for laptops and phones. We compared them in a separate post.

Vision

Groot 19B-A3B

An open 19B mixture-of-experts model (about 3B active parameters) that chats about text and images: attach a photo and ask anything. Apache-2.0.

Groot 19B-A3BGroot Team · Apache-2.0An open 19B mixture-of-experts model that chats about text and images: attach a photo and ask anything.Try it online →

Unlimited OCR 2

Baidu's OCR model turns photos and scans of documents into clean Markdown, keeping tables and formulas. MIT. More in our guide.

Unlimited OCR 2Baidu · MITTurns photos and scans of documents into clean Markdown: text, tables, formulas and layout, page after page.Try it online →

Voice

Paradee 1.2, Phonon 31M and Audio8 TTS

Paradee is an 8.5 MB English text-to-speech model that runs on a CPU; Phonon 31M, the voice model Ollama just added, runs on a CPU too and has character voices; Audio8 TTS Realistic speaks 23 languages and clones voices. Compared with Ming 3.0 TTS in our text-to-speech roundup.

Coming soon

Some models from the same news are not hosted yet, such as Hy-SimulMT (simultaneous translation) and IDOL (3D people from a photo). Their pages list the license and the official links; in each model's “Official models” table, press Request on the sizes you'd like us to host next.

See all models, including those presented in earlier weeks.

LiveWeights is not affiliated with the model authors or with the AI Search channel. Facts come from the official pages linked on each model page.