Gemini Robotics 2, plus Korean labs open 700B-class models

What shipped

Google DeepMind

Gemini Robotics 2 and Gemini Robotics ER 2

Three robot models covering whole-body control, embodied reasoning and on-device use, with ER 2 already open to developers in Google AI Studio.

LG AI Research

K-EXAONE 2.0 750B-A37B

A 750B mixture-of-experts model with 37B active parameters and 256K context, now under Apache 2.0 rather than the earlier custom license.

SK Telecom

A.X K2

A 688B mixture-of-experts model with 33B active parameters and a 262,144-token window, released with Apache 2.0 weights.

Thinking Machines Lab

Inkling-Small

The smaller Inkling is 276B total and 12B active, takes text, image and audio, and ships in BF16 and NVFP4 under Apache 2.0.

Google

Lyria 3.5 in Google Flow Music

Google's music model gets a 3.5 update inside Flow Music, with changes to musicality, lyrics, vocals and creative control.

arXiv

Kimi K3: Open Frontier Intelligence

The technical report for Kimi K3, covering the 2.8T mixture-of-experts design, Kimi Delta Attention and the 1M-token context.

InclusionAI

Ling 3.0 Flash

A 124B model with 5.1B active parameters, 262K context and an MIT license, trained for coding, research and tool use.

Jina AI

Jina Reranker v3.5

A 0.6B reranker with a 131K window for ordering search results, released under a non-commercial CC-BY-NC license.

A quiet week for the big three closed labs, and a loud one for robotics and for anyone who likes downloading very large models.

Gemini Robotics 2

On 30 July Google DeepMind announced Gemini Robotics 2 as an intelligence layer for robots, with three models. Gemini Robotics 2 is a vision-language-action model for controlling full humanoids and bi-arm robots, adding whole-body control and dexterity. Gemini Robotics ER 2 is the embodied reasoning model, which plans multi-step tasks and coordinates teams of robots. Gemini Robotics On-Device 2 runs locally and adapts to a new robot body in a few hours.

Access is staged. ER 2 is available in Google AI Studio and in private preview on Google's enterprise agent platform. The other two go to early-access partners. DeepMind also introduced ASIMOV-Agentic, a new safety benchmark for this kind of system.

The split between a planner that reasons and a controller that moves is what I'd pay attention to. It mirrors how I already organise agents: one that decides, others that do the work. Splitting the roles also makes each piece easier to test on its own, which matters when the thing being controlled is a physical machine with people standing near it. I'd expect most serious robotics stacks to end up looking like this, with a slow thinker on top and a fast executor underneath.

Open weights keep getting bigger

LG AI Research released K-EXAONE 2.0 on 29 July: 750B parameters, 37B active, 256K context, ten languages, and an Apache 2.0 license. The first K-EXAONE used a custom license, so that change matters. SK Telecom released A.X K2 the same day with 688B total and 33B active parameters. Thinking Machines followed its Inkling launch with Inkling-Small on 30 July, and InclusionAI released Ling 3.0 Flash on 2 August under MIT.

Smaller items

The rest of the list was a mix of media, search and research.

Google rolled Lyria 3.5 into Flow Music on 29 July. Jina shipped a 0.6B reranker, though the non-commercial license limits where you can use it. The Kimi K3 technical report appeared on arXiv, which is the document to read if you want to know what Kimi Delta Attention actually does.

What I'd watch next

Whether the robotics models reach anyone outside early-access partners, and whether the Korean labs keep releasing at this size under a clean license.