Gemini 4 Argon lands, Run:ai goes open source

Google shipped a new frontier model, and almost every number in the launch post comes from Google. Below that, the more useful releases for people who run their own stack.

Gemini 4 Argon

Gemini 4 Argon is available today through the API in AI Studio, in Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search for Ultra and Pro. The launch price is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off.

Google says Argon ranks first on the Vals Index, which covers finance, coding, legal and tax work. The examples are the interesting part. Google reports a Chrome refactor that separated core page handling from a roughly 6 million line folder, finished in weeks instead of months. It also reports that the Agent Development Kit team let Argon run overnight, cutting in-memory session prep time by 40% and speeding up cold imports by 21%. And it says Argon sped up xTB, a quantum chemistry package, by up to 2.3x on structure prediction.

All of that is self-reported, run by the people who built the model. The 2.85x speedup on Gemma 4 E2B comes with upstream pull requests, so that one I can check.

The robotics post adds a score: 87% on ERQA, which tests physical and spatial reasoning. Google says that beats every other frontier model. The post is mostly prompts, including a Spot quadruped practicing a foot-pedal bin in a MuJoCo twin before trying it for real.

Liquid AI, on-device

LFM2.5-24B-A2B is a reasoning-only model with 24B total and 2B active parameters, a 131K context window and 128K vocabulary. Liquid AI reports 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B, and 184 tokens per second on an Apple M5 Max. Those are the team's own runs, so treat them as a starting point.

NVIDIA opens Run:ai

NVIDIA released the full Run:ai platform as open source. Four pieces can be adopted separately: KAI Scheduler for GPU-aware Kubernetes scheduling, a GPU Fractioning Operator, KAI Resource Management for quotas, and Workload Map. Crusoe, Nebius, Mirantis and five other partners are integrating it. For anyone sharing a GPU pool across teams, being able to read the scheduler matters.

Smaller items

ComfyUI now has NVIDIA's RTX video effects as local nodes: super resolution up to 4x, frame generation from 2x to 16x, and SDR-to-HDR conversion. Ai2 describes Molmo2Fish, an experimental model built on open Molmo 2 that tracks salmon in sonar footage and corrects errors from plain-language instructions.

What I'd watch next

Whether anyone outside Google reproduces the Vals Index ranking or the ERQA score. Whether LFM2.5 holds its agent numbers on tasks Liquid AI did not pick. And which of the Run:ai partners ships something built on the open scheduler first.