# A quiet week: Qwen3.8-27B, DeepSeek adds vision, DeepMind on games

> Week of 17 August 2026: Qwen3.8-27B opens under Apache 2.0, DeepSeek tests image input on V4 Flash, and a 15-year look back at game-playing agents.

Published 13 Oct 2026. Coverage: 17 Aug 2026 to 23 Aug 2026. By Ali Asif.

## Items

1. **Qwen3.8-27B** (Alibaba, Open weights): A dense 27B vision-language model under Apache 2.0, with a 1M-token hosted context and gains in coding and office tasks over Qwen3.6-27B. [Primary source](https://huggingface.co/Qwen/Qwen3.8-27B)
2. **DeepSeek V4 Flash Vision Exp** (DeepSeek, Frontier): An experimental image-capable version of V4 Flash that matches its text quality, billed at V4 Flash text rates with images costing up to 384 tokens each. [Primary source](https://api-docs.deepseek.com/news/news260821)
3. **From Atari to EVE Online: 15 years of AI research in games** (Google DeepMind, Agents and coding): Traces DQN through AlphaGo and SIMA 2, and announces a partnership to study long-term memory and multi-agent behavior inside EVE Online. [Primary source](https://deepmind.google/blog/from-atari-to-eve-online-building-on-15-years-of-ai-research-in-games/)
4. **StateM: 95.3% on Terminal-Bench 2.1 through scaffold design** (arXiv, Agents and coding): Shows that long-horizon agents fail from losing track of state, not from weak reasoning, and reports a $15 frontier run on the benchmark. [Primary source](https://arxiv.org/abs/2608.15089)
5. **Zetta: a closed-loop embodied agent scaffold** (arXiv, Robotics): Proposes closed-loop learning for physical agents, where most existing scaffolds follow fixed open-loop plans. [Primary source](https://arxiv.org/abs/2608.16590)
6. **SemComp-Bench** (arXiv, Image, video, world): A benchmark that scores video generators on whether the intended outcome actually happens in the clip, not only on how it looks. [Primary source](https://arxiv.org/abs/2608.17426)

Not much shipped this week, which gave me room to read. Three launches, one lab retrospective, and a cluster of papers about how agents fail.

## Qwen3.8-27B

Alibaba released Qwen3.8-27B on 17 August. It is a dense 27B vision-language model, built on Qwen3.6-27B, with improvements in coding and office productivity. The weights are on Hugging Face under Apache 2.0, and the hosted API supports a 1M-token context, reasoning, function calling and structured outputs. A dense model in this size class is what a single workstation GPU can serve, so it's the one I'd pick for a client who needs everything on their own premises.

## DeepSeek adds eyes to V4 Flash

On 21 August DeepSeek published an experimental vision version of V4 Flash. It matches V4 Flash on text and adds image understanding for agent work. Each image is billed as up to 384 input tokens at the normal V4 Flash text rate. It is an experimental public API and the weights are not released. The pricing detail matters more than it looks: a document-reading agent that processes scans could now do it without a separate OCR step.

## Games, as a record of how agents grew up

DeepMind published a long look back at its game research on 21 August, from DQN's Atari results through AlphaGo, AlphaZero, MuZero and AlphaStar to SIMA and SIMA 2, agents that play 3D games from the screen with a keyboard and mouse. The new part is a partnership with Fenris Creations to study continual learning, long-term memory and multi-agent behavior inside EVE Online, starting with an offline copy.

## The papers that caught my eye

StateM argues that long-horizon agents fail because they lose track of mutable state, skip procedures they already know or stop too early. It reports 95.3% accuracy on Terminal-Bench 2.1 at a $15 frontier-model run. Zetta takes a similar idea into robotics, with a closed-loop agent scaffold for physical tasks. SemComp-Bench asks video generators a harder question: did the thing you asked for actually happen in the clip?

## What I'd watch next

Whether Anthropic or OpenAI answer the vision-at-flat-token-price idea, and whether StateM's scaffold results reproduce on other benchmarks.

## My take

StateM is the paper I'd hand to anyone running agents for real work. If the failure mode is losing track of state in a long task, then the fix is in the scaffolding around the model, not in waiting for a bigger one. That matches what I see across my own Claude Code agents: when one drifts on a long bid review, it is nearly always a memory or checklist problem, not an intelligence problem.

## Sources

- https://huggingface.co/Qwen/Qwen3.8-27B
- https://api-docs.deepseek.com/news/news260821
- https://deepmind.google/blog/from-atari-to-eve-online-building-on-15-years-of-ai-research-in-games/
- https://arxiv.org/abs/2608.15089
- https://arxiv.org/abs/2608.16590
- https://arxiv.org/abs/2608.17426

Canonical: https://heyaliasif.com/ai/2026-08-17-week-in-ai