# Grok 4.6, Gemini 3.7 Flash and GLM-5.3 arrive in four days

> Week of 10 August 2026: xAI, Google and Z.ai each ship a coding-focused model, Meta opens Muse Glimmer, and NVIDIA releases Nemotron 3.5 Lightning.

Published 13 Oct 2026. Coverage: 10 Aug 2026 to 16 Aug 2026. By Ali Asif.

## Items

1. **Grok 4.6** (xAI, Frontier): Builds on Grok 4.5 with a focus on long-running agents and interactive, visual work, with a 500K window and effort settings up to xhigh. [Primary source](https://x.ai/news/grok-4-6)
2. **Gemini 3.7 Flash** (Google DeepMind, Frontier): Google's generally available workhorse for coding and agents, with a 1M-token window, 64K output and thinking levels from low to high. [Primary source](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/)
3. **GLM-5.3** (Z.ai, Frontier): A coding-first post-training of the GLM-5.2 base for long-horizon software work, with open weights promised after safety hardening. [Primary source](https://z.ai/blog/glm-5.3)
4. **Muse Glimmer 30B** (Meta, Open weights): A 30B open-weight multimodal agent model under Apache 2.0, distilled from Muse Spark and sized for consumer GPUs. [Primary source](https://huggingface.co/meta-models/Muse-Glimmer-30B)
5. **Nemotron 3.5 Lightning** (NVIDIA, Open weights): A 30B hybrid Mamba-Transformer with only 3B active parameters and up to 1M-token context, open weights in several precisions. [Primary source](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16)
6. **Ling-3.0-tiny** (InclusionAI, Open weights): A 7.9B hybrid-reasoning model with 1.3B active parameters under MIT, validated for local use on DGX Spark and Apple Silicon. [Primary source](https://huggingface.co/inclusionAI/Ling-3.0-tiny)
7. **SL2T sign-language-to-text** (Google DeepMind, Science): A sign-to-text model that runs on pose landmarks on the device, so no raw video leaves the phone, starting with American Sign Language. [Primary source](https://deepmind.google/blog/putting-sign-language-ai-into-users-hands/)
8. **BDH-CQ: In-Context Learning with Recurrent Latent Reasoning** (arXiv, Science): The most upvoted paper on Hugging Face that week, proposing a reasoning model whose recurrent memory updates as inputs arrive at inference. [Primary source](https://arxiv.org/abs/2608.09888)

The release calendar looked like a relay race this week. Three labs handed over a new coding model within four days of each other.

## Grok 4.6

xAI put out Grok 4.6 on 12 August. The announcement says it builds on Grok 4.5 and focuses on long-running agents and more ambitious interactive and visual work, the kind of job where a model stays on task across many steps: researching a topic, working through a codebase, turning an idea into a finished app. The API lists text and image input, a 500K context and reasoning effort from low to xhigh. It is also offered inside Cursor and Grok Build.

## Gemini 3.7 Flash and GLM-5.3

Google released Gemini 3.7 Flash on 13 August and calls it its most intelligent workhorse model. It takes text, image, video, audio and PDF, has a 1M-token window and 64K output, and supports thinking at low, medium and high.

Z.ai followed on 14 August with GLM-5.3. It is a post-training update on the same GLM-5.2 base, tuned for complex software engineering and long-horizon coding. The weights are not out yet: Z.ai says open weights come after safety hardening, so I'd treat it as an API model for now.

## Open weights, the other end of the size range

Meta released Muse Glimmer on 10 August, a 30B open-weight multimodal agent model under Apache 2.0, distilled from its proprietary Muse Spark and sized for a consumer GPU. NVIDIA released Nemotron 3.5 Lightning on 11 August: 30B total, 3B active, a hybrid Mamba-Transformer design, with context listed up to 1M tokens. InclusionAI added Ling-3.0-tiny, a 7.9B model with 1.3B active that runs on a DGX Spark or a Mac.

## Sign language on the phone

On 12 August DeepMind described SL2T, a sign-language-to-text model now behind dictation in Gboard and Live Transcribe on Pixel 11, starting with American Sign Language. It works from pose landmarks captured on the device rather than raw video, which is a sensible privacy choice. DeepMind reports training on more than 100,000 hours across over 50 sign languages.

The paper that drew the most attention was BDH-CQ, a reasoning model with a recurrent memory that updates as it reads its inputs.

## What I'd watch next

Whether Z.ai actually opens the GLM-5.3 weights on the timeline it hints at.

## My take

Three labs shipped coding and agent models in four days, and none of them changes my stack by itself. What matters is the price curve: the cheap Flash tier is now good enough for most of what my agents do, and I only call a flagship for bid reviews and anything with money attached. The sign-language model is my favourite item here, because on-device processing is the privacy pattern I'd want for any Saudi client data.

## Sources

- https://x.ai/news/grok-4-6
- https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/
- https://z.ai/blog/glm-5.3
- https://huggingface.co/meta-models/Muse-Glimmer-30B
- https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
- https://huggingface.co/inclusionAI/Ling-3.0-tiny
- https://deepmind.google/blog/putting-sign-language-ai-into-users-hands/
- https://arxiv.org/abs/2608.09888

Canonical: https://heyaliasif.com/ai/2026-08-10-week-in-ai