Putting the 'open' back in OpenAI with GPT-OSS
Plus: An espionage campaign in open source, and much more.
Hi folks,
The big story this week, as far as Forkable is concerned at least, is OpenAI’s return to open weights some six years after the company abandoned its roots for an entirely proprietary approach.
Elsewhere, Sonatype uncovers a global espionage campaign in open source, while Roblox launches an open source AI system to protect kids from predators. And there is much, much more.
As usual, feel free to reach out to me with any questions, tips, or suggestions: forkable[at]pm.me.
Paul
Open issue
The return of ‘Open’AI
It’s been six years, but OpenAI dropped a pair of open-weight language models this week, marking the company’s first meaningful dalliance with “open” since GPT-2 in 2019.
OpenAI has been promising a return to the open realm for several months now, and the ChatGPT hit-maker finally made its move on Tuesday with the introduction of gpt-oss-120b and gpt-oss-20b.
The two new models have been built with mixture-of-experts (MoE) architecture that selectively activates a small fraction of the model’s parameters during inference. In real terms, this means that at any given time, only a few of the model’s specialized “experts” are actually doing the work, which reduces the compute cost and latency while letting the model retain the breadth of knowledge and capabilities you’d expect from something much larger.
For context, OpenAI also debuted its flagship GPT-5 LLM this week, and while the company hasn’t announced its parameter count, estimates peg the figure at somewhere in the “many trillions” realm. At any rate, GPT-5 and its predecessors are OpenAI’s highly-optimized “closed-weight systems” trained on enormous datasets, with significant fine-tuning for safety, reasoning, and multi-modal abilities.
The new gpt-oss brand of LLM are still fairly capable and competitive with OpenAI’s smaller proprietary models (e.g. o4-mini / o3-mini), but they’re not state-of-the-art in the same way as GPT-5 is. They’re mostly about giving developers, researchers, and companies the ability to run models locally, fine-tune and adapt them for specific domains, and experiment with architecture without starting from scratch.
It’s worth noting that OpenAI is specifically calling these new models “open weight” as opposed to “open source.” Indeed, gpt-oss-120b and gpt-oss-20b are available under an Apache 2.0 licence, but that only refers to the model weights (i.e. the trained numerical parameters that define the model’s behavior) and architecture, which are free to use, modify, and distribute. The full training data, code, or processes behind them, aren’t public. In other words, you can build on what OpenAI has released, but you can’t fully retrace or reproduce how these models were trained from scratch.
Still, it’s a fairly big deal. And it’s a signal that efforts from other companies that have pushed a more “open” LLM ethos, such as Meta and China’s DeepSeek, have forced OpenAI to loosen its grip to stay competitive.
Read more: Introducing gpt-oss [OpenAI blog]
Patch notes
An espionage campaign in open source
Security firm Sonatype revealed that North Korea’s Lazarus Group covertly inserted malware into 234 packages on NPM and PyPI during the first half of 2025, potentially impacting up to 36,000 developers.
Read more: Sonatype uncovers global espionage campaign in open source ecosystems [Sonatype blog]
Protecting kids from predators
Online gaming platform Roblox unveiled Sentinel, an open source AI tool designed to spot predatory language across conversations rather than individual messages. Roblox says that it has already flagged 1,200 cases for the National Center for Missing & Exploited Children in H1 2025
Read more: Roblox rolls out open source AI system to protect kids from predators in chats [AP]
Huawei taps open source to challenge Nvidia
Huawei plans to open source CANN, the toolkit behind its Ascend AI chips, aiming to accelerate developer adoption and foster AI ecosystem independence in China.
In short, it’s designed to counter Nvidia’s CUDA platform, which dominates AI development but remains closed and U.S.-controlled.
Read more: Huawei to open-source AI chip toolkit to challenge Nvidia [TechInAsia]
Microsoft to open source Windows UI framework
Microsoft laid out a multistep roadmap to open source WinUI, a core framework powering Windows 11 interfaces. While no firm timelines were announced, the phased plan includes documentation, community contributions, and migrating development to GitHub.
Read more: Microsoft to open source WinUI components for Windows 11 [WebProNews]
Anaconda hits $1.5B valuation in Series C funding
Anaconda, a company built atop the open source Python and R distribution that helps data scientists build, test, and deploy data-centric projects, announced that it has raised over $150 million in Series C funding. A per a Reuters report, this gives the comapny a valuation of around $1.5 billion.
Read more: Anaconda raises over $150M in Series C funding to power AI for the Enterprise [Press release]
Cline claws in $32M
Cline, developer of a widely used open source AI coding assistant, raised $32 million in combined seed and series A rounds. The funding supports the launch of Cline Teams, an enterprise-focused platform for agentic AI-enabled software development.
Read more: Cline raises $32M in seed and series A funding [Press release]
A new open source LLM from Switzerland
Switzerland’s federal tech institutes — ETH Zurich and EPFL — are building a fully open LLM on Switzerland's Alps supercomputer. The model is trained for multilingual prowess in over 1,500 languages, and is set for public release later in 2025.
This is separate to OpenEuroLLM, a collaborative European effort to develop and share large language models tailored to EU languages and legal frameworks.
Read more: Swiss to launch open source Large Language Model [Le News]
Proton’s open source authenticator
ProtonMail parent company Proton released Proton Authenticator, an open source two-factor authentication (2FA) app for mobile and desktop. It offers encrypted backups, easy imports from other 2FA tools, and no ads or tracking, in alignment with Proton’s privacy-centric ethos.
Read more: Introducing Proton Authenticator – secure 2FA, your way [Proton blog]
And finally…
Smart glasses that can vibe code… WTF?
Brilliant Labs, the company behind the open source, AI-enabled smart glasses dubbed Frame, is back with Halo — an upgrade to its predecessor that packs an AI agent that can see, listen, and speak.
Sporting a micro‑OLED display, bone‑conduction speakers, sensors, and a low‑power AI chip, Halo promises a private, always-on AI assistant that talks back in real time and can even help you build apps via an on-device “vibe mode”. Yup, just verbalize your idea, and Noa — the built-in AI assistant — will spin it up for you.
As with Frame, Halo is entirely open source — every design file, code, and hardware specs are available to the public.
In terms of price, Halo specs will set you back a reasonble-sounding $299, with pre-orders open now. Shipping is expected in Q4, 2025.
Read more: Halo’s open-source glasses comes with AI agent [Design Boom]