Unhinged as Grok may be, it’s now open source. xAI’s CEO, Elon Musk, posted on X that the company made the older Grok 2.5 model available to the public and will do the same with the upcoming Grok 3. For now, anyone can download, run and even tweak Grok, whose source code was uploaded to the Hugging Face platform. However, there are restrictions to xAI’s open-source license, which doesn’t let people use Grok to train, create or improve other AI models.

It’s not the first time xAI has made its models available to the public. In March 2024, the company released the raw base model of Grok-1, which isn’t finetuned for any specific task. As xAI continues to make Grok more accessible, it’s a stark contrast to OpenAI, which has only offered less powerful models of its ChatGPT model to researchers and businesses.

Making Grok open source allows independent developers to potentially improve on the AI model, but xAI is still trying to move past an extremely alarming episode of Grok providing antisemitic responses and referencing itself as MechaHitler. The Grok team attributed the incident to “deprecated code” that has since been fixed. As for Grok 3, Musk also said on X that it will also go open source in six months, but we may have to take that estimated release with a grain of salt, considering the CEO’s other promised timelines.

But OpenAI’s earlier decision to release open-source versions of its powerful model—the first time it’s done so since 2020, may be more consequential. OpenAI’s move follows a flood of Chinese AI models spurred by the surprise release from Chinese AI startup DeepSeek.

It’s a major shift for the U.S. AI developer, now worth $300 billion. Open weight models allow developers to fine-tune for specific tasks without retraining it from scratch. Despite its name, OpenAI has focused on releasing closed, proprietary models, meaning developers couldn’t get under the hood to see how they worked—allowing OpenAI to charge for access to its powerful models. 

DeepSeek tested that strategy. The Hangzhou-based start-up made waves by releasing models that matched the performance of products from Western rivals like OpenAI and Anthropic. By making its technology openly accessible, DeepSeek allowed developers around the globe to experience the power of its models firsthand.

Since then, Chinese AI development has exploded, with companies large and small rushing to unveil increasingly advanced models. Most releases are open-source. 

“Globally, AI labs are feeling the heat as open source models are increasingly recognized for their role in democratizing AI development,” Grace Shao, an China-based AI analyst and founder of AI Proem, says. 

U.S. tech stocks have rebounded from the slump triggered by DeepSeek, but the shift to open-source may be more permanent. In March, OpenAI CEO Sam Altman conceded that the developer may have been on the “wrong side of history” by maintaining a closed approach. 

The race is now geopolitically charged. Ahead of releasing the open-source models, Altman said he was “excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit.” Altman’s statement leans into a growing competition over AI–one that developers in the U.S. are worried of losing.

“This plethora of simultaneous open AI models (with published weights and papers about technique) is an ‘idea orgy.’ The collective innovation should easily soar past anything one company can do alone,” Benchmark general partner Bill Gurley wrote on X in late July. “It’s formidable and should easily win over single proprietary players (anywhere in the globe).”

China embraces open-source

Chinese AI firms are now aggressively championing open-source. 

Baidu, once the leader in China’s AI development with its ERNIE model, went open-source a few months ago to catch up with Alibaba and DeepSeek. Kuaishou and Tencent have both released open-source video-generation models. Zhipu AI, Moonshot AI and MiniMax–some of China’s so-called “AI tigers”—have also released open-source models in recent weeks.

Rather than closely guard their breakthroughs, Chinese developers think an open approach will encourage greater innovation and encourage adoption. “When the model is open-source, people naturally want to try it out of curiosity,” Baidu CEO Robin Li told analysts in February, soon after the company unveiled its plans to go open-source

And there’s a business argument too: Alibaba executives, for example, argue that their open-source Qwen models encourage companies and startups to use Alibaba’s cloud computing services. 

Since DeepSeek’s release, Chinese companies have rushed to integrate Chinese AI models into their products, including social media platforms, cars, and even air-conditioners. 

There may also be a psychological element at play. Going open-source lets users around the world see the power of Chinese AI models for themselves, appealing to an up-and-coming tech sector that’s long been denigrated by outsiders as a copycat.

Export controls

China has supported other open-source technologies. Officials back the use of the RISC-V chip design architecture, an open-source alternative to proprietary architectures like ARM and Intel’s x86. RISC-V allows Chinese chip engineers to share best practices and ideas, spurring the growth of the broader sector. 

Beijing seeks to develop a self-sufficient semiconductor sector, in part due to concerns of the U.S.’s control of critical parts of the chip supply chain. The Biden administration’s decision to impose chip controls in 2022 intensified China’s push for domestic innovation. 

China’s embrace of RISC-V has raised eyebrows in Washington. Last year, the House Select Committee on the Chinese Communist Party recommended that U.S. officials study the risks of RISC-V, and reportedly proposed preventing U.S. citizens from aiding China on the open-source architecture. 

Leaders vs. followers

China’s embrace of open-source aligns with the country’s initial position as a runner-up in AI.

“If you’re an OpenAI, an Anthropic, a Google…if you’re really leading, then you have this incredibly valuable asset,” Helen Toner, the director of strategy at Georgetown’s Center for Security and Emerging Technology, said at the Fortune Brainstorm AI Singapore conference in mid-July. “It’s easy to understand why they wouldn’t want to just hand out [their models] for free to their competitors if they’re able to sell access to their closed systems at a premium.”

But for followers, who “can’t compete at the frontier,” releasing an open-source model is a way to show “how advanced you are,” she explained. 

Open-source models also “buy a lot of goodwill,” Toner, who once served on OpenAI’s board, added. “What we’ve seen over the last couple years is how much soft power is available to people who are willing to and organizations that are willing to make their technology available freely,” she explained.

The U.S. may now recognize the “soft power” potential of open-source. “The United States is committed to supporting the development and deployment of open-source and open-weight models,” Michael Kratsios, director of the U.S. Office of Science and Technology Policy, said in South Korea earlier this week

And with OpenAI’s decision, U.S. AI is now perhaps put in a rare position: Following, not leading.

We receive so many pitches for crowdfunding projects. Since crowdfunding is a bit of a gamble, it is always hard to determine just how to write about these projects, considering that many of them may not ever deliver. Cool Crowdfunding is our way of sharing some of the ones that catch our eye and that we think readers should know about. 

Remember, some of the crowdfunding projects from our previous Cool Crowdfunding collections may still be active! Go back and check them out to see if there’s anything you shouldn’t miss.

---

OpenCharge

When EV charger company Enel X closed the servers on US JuiceBox customers, a lot of EV owners came face-to-face with how much control one corporation can have over a critical appliance. 

OpenCharge is a project hoping to free folks from fearing the loss of fundamental features (like software-set power limits) ever again.

Currently in pre-launch.

SigCore UC

This universal I/O controller is designed for makers and engineers who work with analog and digital signals in the real world. An open source tool for test rigs, burn-in racks, and custom control systems. Industrial control systems are often proprietary, expensive, and very limited. I can see a number of these finding a home on a factory floor.

Coming soon.

Downing’s 5th GeN64 Portable Nintendo 64 DIY Kit

More than a few makers got their start in electronics hacking video game consoles. For a time it was an unofficial right of passage to turn retro gaming hardware into a portable system. But now we’re in the distant future. If you want to build a portable Nintendo 64 you no longer need body filler for your case, and you can even pick up (almost) all of the electronics in a kit.

Closes August 14, 2025

Spiker:bit

As David Groom wrote in the Make Things newsletter, SPIKER:bit is Backyard Brains’ latest kit. Designed with beginners in mind, their latest board is for anyone looking to start controlling electronics with their mind, muscles, or heart. It’s been extensively tested, and can be programmed with Micro:bit coding blocks.

Closes on August 21, 2025

Curved Keyboards

Keyboards with concave keywells may be more ergonomic, but require either a custom build (not opposed) or a significant investment. This tilting kit from 3dkeycap is a set of 3D-printed risers for MX keyswitches, making comfort much more accessible.

Closes on August 28, 2025

Bela Gem

We’re always happy to see a project solving engineering problems for artists, and Bela Gem is a powerful one – with a legacy of installed projects and extensive documentation to boot. Bele Gem is powered by PocketBeagle2 and made to speed up programming and development of audio artwork. Though the campaign recently closed, they’re still taking late orders.

MADE

Not a product, but a worthy cause. The Museum of Art and Digital Entertainment exposes students to technological history through hands-on fun. In 15 years they’ve brought back the first MMORPG, helped change copyright law, worked with dozens of institutions, and educated thousands of students in STEAM subjects. And though they run lean, they need support to do it for the long haul. If you want to support the Bay Area’s only video game museum, you can help them out on GiveButter.

Collabora Online, the online office suite based on LibreOffice, just released its big update for the year. Collabora Online now has faster document editing, security upgrades, and plenty of interface improvements, and you can still self-host it on your own server.

You might not have heard of Collabora, but it’s the company that maintains the official enterprise-ready edition of LibreOffice. Collabora Online is a modified version designed to run in a web browser, competing with services like Google Docs/Sheets/Slides and Microsoft 365. It’s mostly built for companies and other organizations to deploy on their own infrastructure, but you can also host it on your own home server with the CODE edition or NextCloud app.

What’s New in Online 25.04

Collabora Online still has a similar design and layout as LibreOffice, but with some additional features and browser-specific functions on top. This release adds support for autotext and custom dictionaries, so you can add save text snippets (like boilerplate text or standard replies) to use later in any document.

If you’re working on a Writer document with other people, comment threads now show replies more clearly and support tagging other people with @ names. It should work more or less the same as Google Docs or Word in Microsoft 365. There’s also a new navigation sidebar to help you jump around sections in a document.

There are some impressive performance improvements in this release, too. Collabora said in a blog post, “we’ve invested a lot of work into improving the performance at the client-side, in the browser. […] By being smarter about when to render documents and batching DOM changes, to incrementally freeing memory, we have greatly improved performance in some cases bringing performance from 11fps [to] 60fps.”

If you work with PDF files on a regular basis, you might appreciate the new support for the PDF 2.0 standard. It has improvements for digital signatures, transparent elements, annotations, and encryption, but not all PDF readers fully support it yet.

The new update also adds some new spreadsheet functions to close the gap with Excel. You can now use XLOOKUP and XMATCH to search and match data, as well as dynamic array functions like FILTER, SORT, SORTBY, UNIQUE, SEQUENCE, and RANDARRAY.

There are many other changes in this release, including “more than 100 accessibility improvements,” faster font formatting, and better support for high-resolution screens. Collabora Online 25.04 is a lot closer to a feature-complete browser-based office suite, though it’s still not intended for mainstream use.

Try Out Collabora Online

Unfortunately, Collabora Online still isn’t as easily accessible or Google Docs or Microsoft 365. You can set it up on a home server, like a NAS or Raspberry Pi, with the CODE edition.

If you already have a NextCloud server running, you can install the official app to edit documents. That gives you an experience much closer to opening documents in your browser from Google Drive or OneDrive, with easy file synchronization across your devices and the cloud.

Source: Collabora Online