New open source AI models worth knowing this October
Flint, Nanbeige4.3-33B, Unlimited OCR 2, Flora and Kimi K3.1 Pro: what each does and which licence it carries.
By ClaudeGPT Si Newsroom · Editorial team
· 3 min read
Five new open source AI models stood out in late October 2026: Recursive's Flint, Nanbeige4.3-33B, Baidu's Unlimited OCR 2, Nvidia's Flora and Moonshot's Kimi K3.1 Pro, whose weights are due by November 4. Strictly speaking these are open-weight models: you can download and run them, and most come under licences such as Apache 2.0 or MIT, but the training data is usually not published.
Here is a quick tour, with each model's job and licence. ClaudeGPT Si, a free AI news site, covered each one when it landed; you will find the full stories in its AI Models section.
At a glance
| Model | Maker | Released | Good for | Licence |
|---|---|---|---|---|
| Flint | Recursive | October 23 | Coding agents on a laptop | Apache 2.0 |
| Nanbeige4.3-33B | Nanbeige | October 24 | Agentic coding, office tasks | Apache-2.0 |
| Unlimited OCR 2 | Baidu | October 27 | Turning long documents into text | MIT |
| Flora | Nvidia | October 26 (our report) | Labelling robot training video | See Nvidia's pages |
| Kimi K3.1 Pro | Moonshot AI | October 23 (API) | General work, long context | Not yet stated |
Flint: a coding model that fits on a laptop
Flint has about 40B parameters but switches on only around 2.5B of them for each token, according to its model card. That design, called mixture of experts, is what lets it run on a laptop. Recursive aimed it at the daily work of a coding agent: reading a codebase, fixing bugs, adding features, writing tests. Read our Flint story for the tools it works with.
Nanbeige4.3-33B: a mid-size agent model
This one is bigger, at about 33B parameters, with a context window of 512K tokens. It handles English and Chinese. The lab calls it the teacher model of its 4.3 series, meaning it will be used to train smaller versions later.
Unlimited OCR 2: whole documents in one go
Baidu's model is small, at about 3B parameters, and does one job. It reads a long document in a single pass instead of page by page, and Baidu says memory use stays nearly flat as the output grows. Good news for anyone with a folder of scanned contracts.
Flora: open models for robot data
Nvidia's Flora is a set of three models that caption, segment and score robot training video. It matters less to the average reader and more to robotics teams, who spend heavily on labelling footage. Our Flora report has the details.
Kimi K3.1 Pro: the big one, almost here
Moonshot calls K3.1 Pro its most capable 3T-class open-weight model to date, with a context window of one million tokens. You can already use it through the Kimi API and apps. The company says the full weights will be released by November 4, and its launch post does not state a licence yet.
Open source or open weight?
Most of the new open source AI models are open weight. You get the trained model and the right to use it under its licence, but not the full recipe. For most people that is enough: you can run the model privately, fine-tune it and build on it. If you need fully open training data, check each model card before you choose.
When is a closed model still the better pick? When you need the very top of the range today and do not want to run hardware. Anthropic's Claude Opus 5.5 and OpenAI's GPT-6.1 Sol are offered as hosted services through their makers' APIs.
Keep the list fresh
New open models keep arriving on Hugging Face. ClaudeGPT Si writes them up with dated facts and links to each model card, and you can follow along through its RSS feed.