Apple builds local/cloud models with Google: Gemma 4 12B, a laptop-sized model with multimodal power

In today’s edition of Data Points, you’ll learn more about:

• The first working vaccine built by AI
• Kimi CLI, Moonshot’s software engineering agent
• The White House’s plans for an OpenAI stake
• OpenJarvis, an open-source agent that learns on-device

But first:

Apple overhauls Siri, partners with Google on new local models

At its WWDC developer conference, Apple is expected to announce an updated Siri personal assistant. The new Siri still primarily uses a voice interface, but controls both Apple and third-party apps, gathering data and using generative AI to compose emails, schedule calendar events, and perform other routine tasks. Siri will also be more conversational, closer to a ChatGPT experience than its established command-and-response structure. Apple partnered with Google, using access to Gemini models to distill smaller, task-specific versions capable of running locally on Apple devices, while also routing some Siri queries to a cloud-hosted Gemini model for more complex tasks. that the company touts as more secure and better personalized than cloud models. The new applications and models are expected to debut this fall with new operating systems and iPhones. (Bloomberg)

Google’s latest open-weights laptop-sized multimodal model

Google released Gemma 4 12B, a 12-billion-parameter model designed to run on consumer laptops with just 16GB of RAM while delivering multimodal capabilities comparable to its larger 26B variant. The key innovation is its encoder-free architecture—instead of using separate modules to process images and audio before feeding them to the language model, Gemma 4 12B processes visual and audio inputs directly through lightweight embedding layers, reducing both memory overhead and latency. This is the first mid-sized model in the Gemma lineup to support native audio inputs. The model comes equipped with Multi-Token Prediction drafters to speed up inference and is released under an Apache 2.0 license with support across popular inference frameworks like Hugging Face Transformers, llama.cpp, vLLM, and others. Google notes that Gemma models have now crossed 150 million downloads across its ecosystem. (Google)

Cambridge researchers test first vaccine derived from AI analysis

Researchers at the University of Cambridge completed early human trials of a vaccine whose key component was designed entirely by artificial intelligence—a first for the field. The team used AI to analyze genetic codes from multiple coronaviruses and create a “super-antigen” that could theoretically protect against an entire family of viruses, including unknown variants and animal-to-human spillover events. The initial trial involved 39 people and assessed safety; a larger 200-person study will measure immune response. Immune impact was described as “modest” so far, but the approach has generated significant interest because it sidesteps the constant cat-and-mouse game of traditional vaccine design, where manufacturers chase each new mutation. The team is already working on AI-designed candidates for seasonal flu, bird flu, and Ebola. (BBC)

Kimi software agent hopes to rival Codex and Claude Code

Moonshot AI released Kimi Code CLI, an open-source terminal agent that reads and edits code, runs shell commands, and plans its next steps based on feedback. Written in TypeScript and distributed via npm, the tool ships with three built-in subagents—coder, explore, and plan—that run in isolated contexts and can work in parallel. The CLI installs via a single script without requiring Node.js, configures Model Context Protocol servers conversationally through `/mcp-config` instead of raw JSON, and supports video input for dropping screen recordings directly into chat. Moonshot requires either Kimi Code OAuth or a Moonshot AI API key for model access, but the CLI itself is MIT-licensed. It joins a crowded field of terminal coding agents from Anthropic, OpenAI, and Google, distinguishing itself mainly through its native subagent orchestration and conversational MCP setup. (MarkTechPost)

U.S. government hopes to invest in OpenAI

The White House and OpenAI are negotiating a government stake in the company, with discussions ongoing for more than a year since Sam Altman first pitched the idea directly to President Trump in early 2025. Under the potential agreement, OpenAI would donate equity to seed a “Public Wealth Fund” modeled on the company’s April policy proposal, a vehicle designed to let citizens share in AI’s financial upside through direct returns. No investment terms have been finalized, but the talks continued this week as Altman met with lawmakers in Washington. OpenAI is valued above $850 billion and plans an IPO as soon as this year, making any government stake potentially enormous. The Trump administration has already taken equity positions in Intel, IBM, and several quantum and critical mineral companies during its second term. (CNBC)

OpenJarvis, an open framework for on-device learning agents

Researchers at Stanford and Lambda Labs released OpenJarvis, an open-source framework that runs AI agents entirely on local hardware without cloud API calls. The system composes any supported model with five swappable primitives—intelligence, engine, agents, tools, and learning—all configured through a single TOML spec file. Open-weight models running through OpenJarvis land within 3.2 percentage points of the best cloud models on benchmarks while cutting marginal API costs by roughly 800× and latency by about 4×. The key technical advance is “LLM-guided spec search,” which uses a cloud model as a teacher during optimization to propose edits across all five primitives, then runs entirely on-device afterward; this recovers 13–32 percentage points of the cloud-local accuracy gap at 7–11× lower cost than single-primitive optimizers. Installation takes one command and about three minutes, with starter presets for common workflows like daily briefings, code assistance, and scheduled monitoring. (MarkTechPost)

Want to know more about what matters in AI right now?

Read the latest issue of The Batch for in-depth analysis of news and research.

Last week, Andrew discussed the White House’s new executive order on AI and the importance of steering away from overregulation while stopping legitimate cybersecurity risks.

“Over the long term, improved vulnerability detection will make software more secure. When bugs are more easily found, the advantage naturally lies with defenders, who can work to patch them. So having software that enables everyone to find vulnerabilities is a good thing — eventually!”

Read Andrew’s letter here.

Other top AI news and research covered in depth:

• Alibaba’s latest proprietary model, Qwen3.7-Max, adds speed and power, challenging U.S. rivals in the AI race.
• WhaleSpotter pairs sensors with AI algorithms to detect marine mammals, showcasing how AI is saving whales.
• An investigation into the gray market for LLM access reveals how middlemen package extra tokens and hijack IDs to resell and distill models.
• Fine-tuning LLMs to expand on summaries can inadvertently strip models of copyright alignment guidelines, unearthing pretraining texts.

A special offer for our community

In case you missed it, DeepLearning.AI launched our first-ever subscription plan for our entire course catalog! As a Pro Member, you’ll immediately enjoy access to:

• Nearly 200 AI short and long courses from Andrew Ng and industry experts
• Labs and quizzes to test your knowledge
• Projects to share with employers
• Certificates to testify to your new skills
• A community to help you advance at the speed of AI

Enroll now to lock in a year of full access for $25 per month paid upfront, or opt for month-to-month payments at just $30 per month. Both payment options begin with a one-week free trial. Explore Pro’s benefits and start building today!

Try Pro Now

Data Points is produced by human editors with AI assistance.