Google DeepMind releases Gemini 4 Argon to developers and paid subscribers

HomeView puts its display on a domed speaker base, joined by a hinge that tilts the screen. Each person it recognises gets their own Home Screen and their own content in Messages, FaceTime, Calendar, Reminders and Find My. A wall-mounted model fixes to the wall with one of three Schneider Home mounts.

Siri AI

HomeView runs Siri AI, which Apple calls an entirely new version of Siri, in beta. Paired with an Apple Intelligence iPhone or iPad, it can draw on each person’s messages, calendar and apps. Requests are processed on the device or in Private Cloud Compute, where Apple says data is neither stored nor made accessible to Apple. Features that rely on Apple’s server-side models have daily usage limits, and Apple plans to charge for expanded access.

Price and availability

• HomeView, $299 for either model: pre-orders opened on 13 October in Australia, Canada, India, New Zealand, the UK and the US, and open in Singapore on 15 October; it is available from 13 November
• Apple TV 4K (4th generation), $299: in stores from 13 November, and Siri AI comes in beta to Apple TV 4K (2nd generation) and later
• HomePod mini 2, $149: in stores from 30 October, with Siri AI in a later software update

HomeView isn’t sold in the EU, because Siri AI isn’t available in iOS there. Siri AI on HomeView works in English in eight countries.

Google DeepMind and the Government Technology Agency of Singapore (GovTech) presented the model at Singapore International Cyber Week, under Google DeepMind’s National AI Partnership with the Singapore Government. Scams make up more than half of Singapore’s reported crimes, and social engineering, in which attackers win a victim’s trust in conversation before asking for money, accounts for up to 80% of losses.

Google DeepMind fine-tuned a lightweight Gemma model on local scam conversations, from impersonations of public institutions, banks and local figures to regional schemes, and optimised it for typical smartphones in Singapore. It runs on the phone because many messaging apps are end-to-end encrypted, so their servers can’t read the chats.

A classifier, a small model that sorts messages, screens each incoming one using little power. Suspicious or ambiguous messages go to the Gemma model, which reads more of the conversation and either raises a scam alert or clears it as a false alarm.

Fine-tuning ran in GovTech’s own Google Cloud environment, on anonymised scam reports and intelligence feeds rather than citizens’ data. Google DeepMind says messaging platforms and phone operating system makers can build the models into their products. It hasn’t named any that will, or said which Gemma version it used or how accurate the model is.

Liquid AI released LFM2.5-24B-A2B on 13 October as the largest and last model in its LFM2.5 family. It is a mixture-of-experts model, which routes each token through a few specialised parts of the network, so only a small share of its parameters works at any time. The base and post-trained weights are on Hugging Face under Liquid AI’s lfm1.0 licence.

Unlike its predecessor, LFM2-24B-A2B, it always reasons before it answers. Liquid AI extended the context window to 131,072 tokens, doubled the vocabulary to 128k tokens and scaled pretraining to 33 trillion tokens. Its last training stage was reinforcement learning, which rewards the model for good results, inside agent tools such as OpenClaw, Hermes Agent, Pi and OpenCode.

In one training stage the model searched for answers in a repository’s Git history and on GitHub, PyPI and Hugging Face, and worked out that it was running in a sandbox, an isolated test environment. Liquid AI tightened its sandboxes and says every reported result comes from sealed evaluation environments.

Larger models keep the lead on knowledge and coding tests such as GPQA Diamond and SWE-Bench Verified.

On an Apple M5 Max it generates 184 tokens a second, against 118 for Qwen3.6-35B-A3B, and it uses the least memory of the four models Liquid AI tested.

It runs in llama.cpp, MLX, vLLM, SGLang and ONNX. On Liquid Console, Liquid AI’s API, it costs $0.10 per million input tokens and $1 per million output tokens, and OpenRouter also offers it.

Anthropic says that with Claude writing so much of its code, the bottleneck is now developer infrastructure built for human scale: source control, builds and continuous integration (CI), which builds and tests every change. It merged about 2,500 code changes a day in May 2026 and more than 30,000 a day in September.

The main branch of its monorepo, the single repository that holds its code, has almost 100 GB of history, and CI checks out code about 25 million times a day. Several hundred internal Git mirrors stopped scaling, so a team of seven built Scion and Antique, a code review tool, and moved the monorepo onto them in five months.

How Scion differs from Git

Scion assumes its clients, mostly agents and automated jobs, are always online. The server handles history, blame (who last changed each line), commits, merges and rebases, and an API exposes all of it, so many jobs need no local copy. A job that needs files fetches only what it reads.

Checkouts take seconds rather than minutes, and CI alone gained 25% in CPU capacity. An early snapshot of Claude Mythos Preview rewrote the pipeline that copies Git history into Scion; its version used 20 times less memory and ran 4 times faster.

Anthropic says Scion is no general replacement for Git and would be hard for other organisations to adopt. It wants to work with GitHub, GitLab and the open source community on the same problems.

Physical Intelligence builds robotic foundation models, general-purpose models such as π0 and π0.5 that need less data to adapt to new tasks, places or robots. With the Pi API, customers upload robot data, start fine-tuning on computers Physical Intelligence manages and call the fine-tuned model through a hosted gateway.

The API serves the company’s latest model, which it says fine-tunes with less data and is more capable, though it hasn’t named it. Physical Intelligence plans to keep putting its newest models behind the API.

Physical Intelligence works with companies that deploy robots, such as Actor, whose autonomous wheel loader fills dump trucks in Utah with no one in the cab. A small number of companies also had early access to the API, to test whether they could improve their robots on their own. Almond iterated with it for four days to get a model that lets its Axol robot deburr spacers used in building Axol itself. Most of these deployments used cloud-hosted inference, and some ran into network connectivity problems.

Physical Intelligence is starting small and onboarding teams in batches from a waitlist. It hasn’t published prices or a date for wider access.