AI2 takes on DeepSeek; Microsoft brings back DocumentDB

Welcome to Runtime! Today on Product Saturday: The Allen Institute for AI releases an actual open-source challenger to DeepSeek's V3 model, Microsoft open-sources a NoSQL database under an old and familiar name, and the quote of the week.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

Ship it

Beast mode: Long before DeepSeek scrambled everyone's circuits, the Allen Institute for AI was quietly toiling away in Seattle working on ways to build world-class large-language models without the resources enjoyed by companies like OpenAI and Anthropic. This week it introduced Tülu 3 405B, a beefed-up version of the Tülu 3 model released last year that "achieves competitive or superior performance to both DeepSeek v3 and GPT-4o," the nonprofit said in a blog post.

The new model is AI2's first with "Reinforcement Learning with Verifiable Rewards (RLVR), a novel method we introduced for training language models on tasks with verifiable outcomes such as mathematical problem-solving and instruction following," it said. And while both AI2 and DeepSeek released the weights behind their models, which helps others tweak them to suit their own needs, AI2 also released the code and training data for Tülu 3 405B.

Reason to believe: Not content to let the week end with DeepSeek still on everyone's mind, on Friday OpenAI released its latest reasoning model, the o3-mini, which it said was "the newest, most cost-efficient model in our reasoning series." OpenAI first showed off this model in December, but Wired reported that the company moved up the launch date after watching buzz build around DeepSeek over the last month.

"While OpenAI o1 remains our broader general knowledge reasoning model, OpenAI o3-mini provides a specialized alternative for technical domains requiring precision and speed," the company said. It's available now for ChatGPT users and for developers via an API, but VentureBeat noted that while o3-mini is certainly one of OpenAI's most affordable models, it "pales in comparison" to DeepSeek's pricing for R1's API.

Run it back: One of Microsoft's most widely used databases is Azure Cosmos DB, a document database designed to run distributed applications that has a long history inside the company. Last week Microsoft announced that it will release a MongoDB-compatible version of that database under the MIT license under the new — yet pretty old — name of DocumentDB.

The new DocumentDB "is intended to provide a standard NoSQL environment for your data to reduce the complexity associated with migrating from one platform to another," according to InfoWorld. In a blog post, Microsoft said it wants to make DocumentDB into an actual standard: "DocumentDB is the first implementation of the project’s more ambitious mission to create a standard for open-source document databases, much like the ANSI (American National Standards Institute) SQL standard for relational databases."

Hug it out: The rise of alternative infrastructure computing providers has been one of the more interesting aspects of the generative AI boom, proving once again that when a new technology comes around, early adopters don't necessarily want to do things the old-fashioned way. Developers on Hugging Face now have four new options for inference computing that they can select directly from its platform: fal, Replicate, Sambanova, and Together AI.

Hugging Face has offered inference options for several years, but "we’ve refined our core value proposition towards collaboration, storage, versioning, and distribution of large datasets and models with the community," it said in a blog post. The four providers offer serverless inference, which means developers don't have to worry about actually configuring the hardware that will run their models.

Master of puppets: After Portland's own Puppet was acquired by Perforce two years ago, the new company announced that all future updates to its infrastructure-as-code product would be closed-source. That didn't sit well with the community of open-source Puppet users who have been using the software for years, and this week they delivered OpenVox, a proper fork of old-school Puppet.

It's a similar playbook followed by users of HashiCorp's Terraform (which led to the release of OpenTofu) and Redis, which led to Valkey. "The group intends to maintain downstream compatibility for as long as possible," according to The New Stack.

Stat of the week

Most CISOs are still fighting for respect and clout within their organizations after years of being seen as a roadblock to shipping software. But things are looking up, according to a new survey from Cisco's Splunk: "More than eight of every 10 CISOs now report directly to company CEOs, compared with less than half just two years ago," according to Cybersecurity Dive.

Quote of the week

"They understand the power and the limits of the models, and I don’t think there’s any chance they just YOLO some model output into a nuclear calculation." — OpenAI Chief Product Officer Kevin Weil, revealing in a Reddit AMA on Friday that he's not paying super close attention to the current direction of the federal government.

The Runtime roundup

Hired goons working for Elon Musk on Friday "locked career civil servants out of computer systems that contain the personal data of millions of federal employees, according to two agency officials," Reuters reported. "Let’s be crystal clear about what’s happening: A private citizen with zero Constitutional authority is effectively seizing control of critical government functions," wrote Mike Masnick at Techdirt.

Thanks for reading — see you Tuesday!