Frameworks and open source
Notable releases in open models, frameworks and tools.
Updated 24 Oct, 7 PM ET
Shanghai AI Lab releases Atria Dawn, an open model for long research tasks (opens the original source in a new tab)Built on GLM-5.2, a 744-billion-parameter model, Atria Dawn targets research and engineering jobs that take many steps and tools, and aims for results others can check. It uses the MIT licence.
Tencent releases WeVisOne-VL, open models that find, count and cut out things in images (opens the original source in a new tab)Built on Qwen3-VL, the 2B and 4B models answer requests about a picture by drawing boxes, pointing, counting, reading text or tracing an object's exact outline. They use the Apache 2.0 licence.
Tencent · Original source (opens in a new tab)
- Tencent releases Hy-Nomas, an open model that builds a video world you can steer (opens the original source in a new tab)
Hy-Nomas makes video one short piece after another, and between pieces you can add objects, change the scene or move the camera with text. The weights are free to use under the Apache 2.0 licence.
Tencent · Original source (opens in a new tab)
Google's ADK 2.13 adds live voice chats and MongoDB memory for agents (opens the original source in a new tab)
The Python kit for building AI agents moves 12 experimental features to stable, adds real-time voice chats that users can interrupt, and can keep sessions and long-term memory in a MongoDB database.
Google · Original source (opens in a new tab)
Ai2 releases OlmoEarth v1.3 and ready-made satellite embeddings of the whole planet (opens the original source in a new tab)The model turns a year of satellite images into 128 numbers, called an embedding, for every 10-metre patch of Earth. Ai2 shares these for the whole planet, plus the weights, training data and code.
Recursive releases Flint, an open coding model that runs on a laptop (opens the original source in a new tab)Recursive says Flint has 40 billion parameters but uses only 2.5 billion for each token (word piece), writes over 125 tokens a second on an M4 Pro MacBook Pro, and was built with its automated research system.
Recursive · Original source (opens in a new tab)
IBM releases Granite Speech 5.0, open models that turn speech into text (opens the original source in a new tab)The new 2-billion-parameter model transcribes six languages, including English, French and Japanese, and translates them to and from English. IBM also updated a faster small model and one that tracks who spoke when.
NVIDIA releases Kumo Relational, a model that predicts straight from business databases (opens the original source in a new tab)
NVIDIA says it forecasts things like customer churn or fraud from a company's tables with no task-specific training, and beat trained models across 41 tasks on 15 databases.
NVIDIA · Original source (opens in a new tab)
DeepSeek releases DeepSeek-V4.1-Flash-1017, an update to its open model (opens the original source in a new tab)
DeepSeek says the new version does better at reasoning, coding and agent tasks, and finishes tasks faster with shorter answers. Anyone can download the weights under the MIT licence.
DeepSeek · Original source (opens in a new tab)
TorchTPU lets PyTorch programs run natively on Google's TPU chips (opens the original source in a new tab)
Google and Meta's PyTorch team built this open-source backend so existing PyTorch code runs on TPUs, Google's AI chips, with small changes such as swapping "cuda" for "tpu".
PyTorch · Original source (opens in a new tab)