# Gemma 4 goes Apache 2.0, Qwen3.6-Plus, GLM-5V-Turbo

> Google opens Gemma 4 under Apache 2.0, Alibaba ships Qwen3.6-Plus with a 1M window, and Z.ai releases a vision-first agent model.

Published 13 Oct 2026. Coverage: 30 Mar 2026 to 5 Apr 2026. By Ali Asif.

## Items

1. **Gemma 4** (Google DeepMind, Open weights): Four sizes (E2B, E4B, a 26B mixture-of-experts and a 31B dense) under Apache 2.0, with the 31B ranked third among open models on the Arena text leaderboard, per Google. [Primary source](https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/)
2. **Qwen3.6-Plus** (Alibaba Qwen, Frontier): A 1M-token context by default and a repository-level coding pitch, with Alibaba reporting 61.6 on Terminal-Bench 2.0 and 78.8 on SWE-bench Verified. [Primary source](https://qwen.ai/blog?id=qwen3.6)
3. **GLM-5V-Turbo** (Z.ai, Agents and coding): Z.ai's repo logs the April 2 release of a multimodal model aimed at vision-based coding and agent work. [Primary source](https://github.com/zai-org/GLM-V)
4. **The Latent Space: Foundation, Evolution, Mechanism, Ability, and Outlook** (arXiv, Science): A survey of latent-space research that sat near the top of Hugging Face's weekly papers. [Primary source](https://arxiv.org/abs/2604.02029)
5. **Generative World Renderer** (arXiv, Image, video, world): A paper on generating rendered visual worlds, also near the top of the weekly list. [Primary source](https://arxiv.org/abs/2604.02329)
6. **Claude Platform release notes, March 30 entry** (Anthropic, Agents and coding): The Message Batches output cap rose to 300k tokens for Opus 4.6 and Sonnet 4.6, and the 1M-context beta for older Sonnets was set to retire April 30. [Primary source](https://platform.claude.com/docs/en/release-notes/overview)

Early April was an open-weights week for me, with one closed launch worth setting next to it.

## Gemma 4 and the license change

Google released Gemma 4 on April 2 in four sizes. E2B and E4B are built for phones and small boards. Then there is a 26B mixture-of-experts model that activates about 3.8B parameters per token, and a 31B dense model. The license is Apache 2.0. Earlier Gemma releases shipped with custom terms, so this switch matters more to me than any chart.

Google says the 31B ranks third among open models on the Arena text leaderboard and the 26B sixth. That is Google's framing. The hardware claim is the useful part: unquantized weights for the 26B and 31B fit on a single 80GB H100, and quantized versions run on consumer GPUs. Context is 128K on the edge models and up to 256K on the bigger two, and training covered more than 140 languages. For a market where tender documents arrive in Arabic and English, that language count is a reason to test it.

I run agents against tender packs and supplier quotes. Those files are the ones I would rather not push through someone else's API. A mid-sized model with a license that lets me ship it inside a product changes what I can tell a client about where their data sits.

## Qwen3.6-Plus

Alibaba's Qwen3.6-Plus also arrived on April 2, with a one-million-token window by default and a pitch built around coding agents that work across a whole repository. The scores above are Alibaba's own table, not mine. It runs through Alibaba Cloud Model Studio and Qwen Chat, and Alibaba says it works with third-party coding tools including Claude Code and Cline. To be clear, this one is not an open-weights release.

## The smaller items

Z.ai's GLM-V repository logs GLM-5V-Turbo on April 2, a multimodal model aimed at vision-based coding and agent tasks. On Anthropic's platform notes, March 30 shows the batch output cap going to 300k tokens for Opus 4.6 and Sonnet 4.6. Two papers led the Hugging Face list: a survey of latent-space research and a generative world renderer.

What I'd watch next: whether anyone publishes an honest Arabic-document comparison of Gemma 4 against the Qwen line, because the leaderboard numbers will not tell me that.

## My take

A 31B model that fits on one 80GB card and ships under Apache 2.0 is the first thing in a while I can picture running next to a tender pack without the pack leaving the building. Saudi buyers ask where documents go before they ask what the model scores. I would want to see how it reads a messy bilingual BOQ before I trust a leaderboard rank.

## Sources

- https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/
- https://qwen.ai/blog?id=qwen3.6
- https://github.com/zai-org/GLM-V
- https://arxiv.org/abs/2604.02029
- https://arxiv.org/abs/2604.02329
- https://platform.claude.com/docs/en/release-notes/overview

Canonical: https://heyaliasif.com/ai/2026-03-30-week-in-ai