# Anthropic's Cowork, DeepSeek's Engram and a new open image editor

> Week of 12 Jan 2026: Claude gets a file-working agent mode, DeepSeek proposes conditional memory, and Google, Zhipu and Black Forest Labs release open.

Published 13 Oct 2026. Coverage: 12 Jan 2026 to 18 Jan 2026. By Ali Asif.

## Items

1. **Claude Cowork (research preview)** (Anthropic, Frontier): Claude Code's agent approach, packaged for non-developers: point it at a folder and it works on the files. [Primary source](https://claude.com/blog/cowork-research-preview)
2. **TranslateGemma** (Google, Open weights): Open translation models built on Gemma, in sizes that fit on ordinary hardware. [Primary source](https://huggingface.co/google/translategemma-4b-it)
3. **MedGemma 1.5** (Google, Open weights): An updated open medical model from Google, small enough to run inside a hospital's own environment. [Primary source](https://huggingface.co/google/medgemma-1.5-4b-it)
4. **GLM-Image** (Z.ai (Zhipu), Image, video, world): Zhipu's open image model, strong on rendering text inside images. [Primary source](https://z.ai/blog/glm-image)
5. **FLUX.2 [klein]** (Black Forest Labs, Image, video, world): Small, fast FLUX.2 variants aimed at near-interactive image generation and editing. [Primary source](https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence)
6. **Conditional Memory via Scalable Lookup (Engram)** (DeepSeek, Science): DeepSeek adds a lookup-based memory as a second axis of sparsity next to mixture-of-experts. [Primary source](https://arxiv.org/abs/2601.07372)

Agents left the terminal this week. That is the thread I would pull on.

## Cowork

On 12 January Anthropic released Cowork as a research preview. If you have used Claude Code, you know the idea: give the model a goal, let it use tools, and let it work through a task over many steps. Cowork brings that to people who never open a terminal. You give Claude access to a folder on your machine and ask for an outcome, such as organising files, pulling figures out of a set of documents, or drafting a report from what is there.

Why it matters is who it reaches. The coding agents of 2025 proved the pattern, but the people who have the most repetitive file work are in finance, procurement and operations. Those are the roles I know best. Cowork is the first mainstream packaging of the pattern aimed at them.

The risk side is real and Anthropic says so in its own post. An agent with file access can delete things or act on instructions hidden inside a document. I would start with a copy of a folder, never the original, and I would read the plan before approving it.

## Open models

Google released TranslateGemma and MedGemma 1.5. Both are small and open, which is the point: a hospital or a ministry can run them where the data lives. Zhipu's GLM-Image targets accurate text in images, and Black Forest Labs' FLUX.2 [klein] pushes image generation toward real-time speeds.

## DeepSeek's Engram

DeepSeek published a paper on conditional memory through scalable lookup, which gives a model a cheap place to store and retrieve static knowledge instead of spending compute re-deriving it. It sits alongside mixture-of-experts as another route to doing more with less. I cannot validate the results, so it goes in the log as a direction to track.

I should be honest about my own use. I run agents against folders of working documents, and the single most valuable habit has been writing down, before I start, what a finished job looks like. An agent with a clear definition of done is a good employee. One without it will happily produce twenty pages that look right. Cowork makes this easier to try, and I expect many people to learn that lesson the same way I did.

**What I'd watch next:** how Cowork handles permissions and prompt injection once people outside the early-adopter crowd start using it.

## My take

Cowork is the one I would have built for myself. Most of my working day is documents: supplier quotes, tender packs, spreadsheets. An agent that lives in a folder and does the tedious part is closer to how an operator works than a chat window is. The caution is permissions. I only let an agent touch a folder after I decide what is in it.

## Sources

- https://claude.com/blog/cowork-research-preview
- https://huggingface.co/google/translategemma-4b-it
- https://huggingface.co/google/medgemma-1.5-4b-it
- https://z.ai/blog/glm-image
- https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence
- https://arxiv.org/abs/2601.07372

Canonical: https://heyaliasif.com/ai/2026-01-12-week-in-ai