NederlandsOpen source and open weights
Open source means the blueprint of software is public and you may use it freely. With an open AI model, the weights, the numbers in which it stores its knowledge, can be downloaded.
With open source, the source code of software is public, and anyone may use, inspect and modify it under certain conditions. Those conditions are set out in a licence. With AI models you mostly come across the term open weights. The weights, or parameters, are the numbers in which a model stores everything it has learned. If they can be downloaded, you can run the model on your own computer.
It is like a recipe that a restaurant shares: with the weights you have the finished dish and can cook with it. With all the sources as well, such as the training data and the training code, you could rebuild the whole recipe yourself.
How does it work?
Open is a spectrum. Some makers share only the weights, others also the code to run or train the model. A licence usually comes with it that determines what is allowed. The large DeepSeek model, for instance, is under the permissive MIT licence, which means anyone may download and use it for free. Nvidia's OpenShell, software that shields AI agents, falls under the Apache 2.0 licence.
What is called open is not always available everywhere. The Chinese tech company Baidu, for example, released a large model that developers can train further themselves, but its licence does not apply in the EU, the United Kingdom and South Korea (link).
Why it matters
Open models can be checked, improved and run on your own hardware by anyone. The company Unsloth, for instance, made a heavily shrunken version of the open DeepSeek model that runs on a powerful laptop or mini computer with 128 gigabytes of memory, although that costs some accuracy (link). The blueprint for shielding agents is open too: Nvidia offers the software for free, so that anyone can use it (link).
For researchers and small companies, that makes the difference: they do not have to depend on the servers of a few large companies, and they can build on what others have already done.
Articles about Open source and open weights
-
Tencent lets AI build its own tools for protein research and mathematics
Tencent has presented the second version of its research AI, Hyra. It delivers not only solutions but also tools that can be used again later. Models for proteins and molecules are already public, so other researchers can now test them themselves.
-
Rocky Linux creator builds Linux-compatible kernel in eleven days with an AI agent
Gregory Kurtzer, known for CentOS and Rocky Linux, built Ginux in eleven days: a new core for an operating system that can run existing Linux programs. An OpenAI AI agent wrote almost all the code. It is still a research prototype.
-
Developer builds a complete processor inside the weights of an AI language model
Developer Divyansh Agrawal built a complete processor inside the weights of the language model Qwen 3.5. In his tests, the model now calculates almost without errors, with no outside tool. It is still a slow experiment, but everything is online for anyone to see.
-
DeepSeek shares code of an AI coding assistant that picks up work on its own
On 31 October, the Chinese AI company DeepSeek released the code of a trial version of an AI assistant that takes on and carries out programming work on its own. Anyone may use the code; it is research code, not yet a finished product.
-
Microsoft releases free AI model that shows how proteins move
With BioEmu-2, researchers can predict which shapes a protein or other biomolecule can take, also together with DNA, RNA or small molecules. Microsoft shares the code and the model for free under an open licence; it is meant for research.
-
AI technique makes weather forecasts for two to six weeks ahead much better
Researchers from Microsoft, MIT, the European weather centre ECMWF and other institutions are teaching a computer to recognise and correct the recurring errors of weather models. The gain is largest around the equator; in Kenya the technique is already being tested.
-
Mistral releases a giant AI model for anyone to use, strong in coding and security
France's Mistral is putting Mistral Large 4, nicknamed Le Chonk, online for free: a model with a trillion parameters that anyone can download and, with enough computers, run themselves. According to the company, it is the best open model for programming and for securing software.
-
Nvidia releases free speech model that transcribes recordings in 40 languages, translates them and recognises speakers
Nvidia has released an open speech model that transcribes, translates and shows who is speaking, all in one go. It uses only a small part of its computing power per step and runs on a single graphics card, without the cloud. For now, the performance figures come from Nvidia itself.
-
AI agents turn old console games into PC versions, but the volunteer community is struggling with it
With smarter AI models and the free toolkit Reverse Engineer Anything, dozens of unofficial PC versions of console games are appearing in a short time. The site that tracked this work is closing its leaderboard: according to its creator, the fast AI work is hurting the community.
-
AI agents turn the longest mathematical proof ever into checkable code in 24 days
The American company Axiom Math had dozens of AI agents turn the proof of 'the enormous theorem' about symmetry, some 15,000 pages, into code a computer can check. Along the way they found 34 deviations in the published work, which have now been repaired.
-
AI from Google and NASA finds many more methane leaks in satellite images
Google Research and NASA built an AI model that detects methane clouds in measurements from the International Space Station. In a test it found about 50% more likely methane plumes than human analysts. The model, code and map are freely available.
-
Free tools take the brakes off Nvidia graphics cards: up to twice as much AI computing power
Free open-source tools remove a brake that Nvidia built into the firmware of its GeForce graphics cards. For AI calculations, some cards become up to twice as fast. In practice the gain is smaller, and the tools are still experimental.
-
OpenAI will now train its most powerful AI models without internet access
OpenAI no longer lets its strongest AI models onto the real internet during training. From now on they practise with a stored copy of the web. At OpenAI's request, the models themselves searched for holes in this new shielding, so the company could close them.
-
Chinese AI model Kimi K3.1 Pro ranks fourth in independent leaderboard
Kimi K3.1 Pro from the Chinese company Moonshot AI ranks fourth in the independent leaderboard of Artificial Analysis, ahead of the newest models from OpenAI and Google. It improved most in independent computer work. Moonshot plans to release the model by 4 November.
-
Hugging Face releases open AI model that lets researchers 'talk' to DNA
With Hugging Face's Carbon-X, researchers paste a piece of DNA into a chat and ask in plain language whether a change is likely to be harmful. The model is free and runs on a researcher's own computer. It is meant for research; the next step is testing in the lab.
