# Gemini Omni 1.1 Flash video, Tencent's Hy4 and GLM-5.3-Flash open up

> Week of 24 August 2026: Google ships a video model with native audio, Tencent opens a 770B Hy4 preview, and Z.ai releases GLM-5.3-Flash under MIT.

Published 13 Oct 2026. Coverage: 24 Aug 2026 to 30 Aug 2026. By Ali Asif.

## Items

1. **Gemini Omni 1.1 Flash** (Google DeepMind, Image, video, world): A generally available video generation and editing model with native audio, output up to 4K, scene extension and video references. [Primary source](https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-omni-1-1-flash/)
2. **Hy4 preview** (Tencent, Open weights): A 770B mixture-of-experts model with 49B active parameters and over 1M tokens of context, open-sourced with an Apache 2.0 license. [Primary source](https://huggingface.co/tencent/Hy4-preview)
3. **GLM-5.3-Flash** (Z.ai, Open weights): The first natively multimodal model in the GLM-5 series: 320B total, 18B active, 1M context, MIT license and ungated weights. [Primary source](https://huggingface.co/zai-org/GLM-5.3-Flash)
4. **Qwen3.8-Flash-Next** (Alibaba, Open weights): An experimental open preview of the architecture planned for Qwen4, with 125B total and 6B active parameters. [Primary source](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)
5. **Granite 4.2 30B** (IBM Research, Open weights): IBM's Apache 2.0 reasoning model with 128K context and tool calling, released with 8B and 3B siblings. [Primary source](https://huggingface.co/ibm-granite/granite-4.2-30b)
6. **Gemini 3.5 Transcribe** (Google, Frontier): A speech-to-text model from the Gemini line, released as its own product for transcription work. [Primary source](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/)
7. **Piloting the world's first double-blind AI evaluations** (Google DeepMind, Policy and business): Outside evaluators tested a Gemini model without seeing its weights, and Google never saw the test prompts, which limits benchmark contamination. [Primary source](https://deepmind.google/blog/piloting-the-worlds-first-double-blind-ai-evaluations/)
8. **Apodex 1.1: Scaling Agentic Intelligence for Complex Work** (arXiv, Agents and coding): A paper on models that work with files, executable code and state across long tasks, including recovering from their own failures. [Primary source](https://arxiv.org/abs/2608.23283)

The week ended with a cluster of open models from China and a video model from Google. Underneath, the item I thought about most was about testing, not building.

## Gemini Omni 1.1 Flash

Google released Gemini Omni 1.1 Flash on 27 August as a generally available model for video generation and editing. It works through the Gemini API and takes text, image and video input. It outputs video with native audio at 360p, 720p, 1080p or 4K, in clips of 3 to 10 seconds at 24 frames per second. Scene extension, interpolation, upscaling and video references are supported.

For marketing teams, the change that matters is conversational editing: you ask for a change to an existing clip instead of regenerating from scratch. I make short videos for my own channel, and re-rolling a whole clip for one fix is where the time goes.

## Three open releases

Tencent open-sourced Hy4 preview on 28 August: 770B total parameters, 49B active, over 1M tokens of context, Apache 2.0, aimed at coding, office work and scientific research. Z.ai released GLM-5.3-Flash on 26 August, its first natively multimodal GLM model, with image, video and file input and ungated MIT weights. Alibaba posted Qwen3.8-Flash-Next the same day as an experimental preview of the architecture it plans for Qwen4.

IBM added Granite 4.2 on 25 August in 30B, 8B and 3B sizes. The 30B is a dense reasoning model with a 128K window, tool calling and an Apache 2.0 license.

## Gemini 3.5 Transcribe

Google also released a dedicated transcription model on 26 August. If you work with Arabic-English meeting audio, as I do, a new competitor in speech-to-text is worth a test on your own recordings.

## Double-blind evaluation

On 27 August DeepMind described a pilot of the first double-blind evaluation of a proprietary frontier-class model. It ran with the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons, using a Gemini Flash Lite model and confidential benchmarks inside Google Cloud's Confidential Space. The evaluators never touched the weights, and Google never saw the test prompts.

## What I'd watch next

Whether other labs agree to blind evaluation, and whether any regulator in the Gulf starts asking for it.

## My take

The double-blind evaluation is the item that matters to my line of work. I came up in audit, where you don't let the party being tested see the test, and that is exactly what Google did here with outside evaluators and confidential benchmarks. If Saudi regulators or giga-project clients ever ask me to prove that an AI system does what I say, this is the kind of method I'd want to point to.

## Sources

- https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-omni-1-1-flash/
- https://huggingface.co/tencent/Hy4-preview
- https://huggingface.co/zai-org/GLM-5.3-Flash
- https://huggingface.co/Qwen/Qwen3.8-Flash-Next
- https://huggingface.co/ibm-granite/granite-4.2-30b
- https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/
- https://deepmind.google/blog/piloting-the-worlds-first-double-blind-ai-evaluations/
- https://arxiv.org/abs/2608.23283

Canonical: https://heyaliasif.com/ai/2026-08-24-week-in-ai