Gemini 3.8 Live, open Qwen-Image-2.1 and Anthropic's Accenture pact

What shipped

Google DeepMind

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking

Two real-time voice models, one built for scale across 97 languages and one for multi-step work, both reachable through the Gemini API.

StepFun

Step 5 Preview

A 600B mixture-of-experts model with 27B active parameters and a 1M-token window, aimed at agentic software and finance work.

Alibaba

Qwen-Image-2.1

An open 7B image generation and editing model with native transparency and up to ten reference images for identity-preserving edits.

PrismML

Ternary Bonsai 2 27B

A ternary model about nine times smaller than full precision that PrismML says keeps 98.2% of its benchmark performance.

Cactus Compute

Needle 3

A sliceable 25M to 121M parameter tool-calling model where every depth from 2 to 20 layers is its own deployable size.

xAI

Grok Voice Transcribe 2.0

A speech-to-text model xAI says is twice as accurate as version 1.0 at the same price.

Anthropic

Partnering with Accenture on embedded evaluation

Each company expects to invest at least $1 billion over five years in outside red-teaming and safeguard testing of frontier models.

arXiv

Vidu S2: Real-Time Interactive, Editable, and Spatial Video Generation

A real-time digital-character model and a real-time video editing model, with early work on spatial video generation.

Voice, images and one policy story that I think will outlast the model news.

Gemini 3.8 Live

On 15 September Google announced two live voice models. Gemini 3.8 Live is designed to be economical and scalable, handles visual input and 97 languages, and is rolling out to everyone in Search Live. Gemini 3.8 Live Extended Thinking is the heavier one for multi-step work. Google says it ranks first on the Artificial Analysis Speech to Speech Quality Index with 82.6 and scores 97.7% on Big Bench Audio. Developers get both through the Gemini API and Google AI Studio, and enterprises can join a private preview.

Those rankings come from Google's own announcement, so I'd run my own audio before trusting them. Arabic dialect handling is the test that matters to me.

Step 5 Preview

StepFun released Step 5 Preview on 20 September, a 600B mixture-of-experts model with 27B active parameters. It takes text, image and video, supports a 1M-token window with up to 64K output, and StepFun points to finance as a particular strength. It is API-only at launch.

Open images and tiny agents

Alibaba's Qwen team opened Qwen-Image-2.1 on 20 September: a 7B generation and editing model with native transparent backgrounds, and up to ten reference images with local edit control. PrismML announced Ternary Bonsai 2 27B on 17 September, based on Qwen3.8 27B, about nine times smaller on disk than full precision and holding 98.2% of aggregate benchmark performance on its own suite (83.9 against 85.4). Cactus Compute released Needle 3, a tool-calling model that you can slice to anywhere from 2 to 20 layers.

xAI released Grok Voice Transcribe 2.0 on 18 September, saying it is twice as accurate as the first version at the same price.

The Accenture agreement

On 18 September Anthropic announced a partnership with Accenture, led by Accenture's Faculty business, for independent embedded evaluation of its frontier models. That covers red-teaming, alignment assessments and safeguard testing. Each company expects to invest at least $1 billion over five years, and Anthropic funds the work directly. It is non-exclusive, and Anthropic is also piloting with METR. The announcement admits that standards for evaluators' access and reporting do not exist yet.

A paper to skim: Vidu S2 on real-time editable video.

What I'd watch next

Who else signs up to fund outside evaluators, and whether anyone writes the access and reporting standards that the announcement says are missing.

What shipped

Gemini 3.8 Live

Step 5 Preview

Open images and tiny agents

The Accenture agreement

What I'd watch next

Sources