Analysis

Chinese AI is not one race: four releases worth separating

DeepSeek, Qwen, Kimi, and GLM are competing across different tasks. Treating them as interchangeable hides the useful differences.

By Saud AlotaibiPublished 3 October 20263 min readSources checked 3 October 2026
The takeaway

Compare exact releases and tasks, rather than ranking entire companies under one label.

“Chinese models are catching up” is a headline that conceals several different stories. A coding model, an image editor, and a multimodal API are not answering the same question. A useful reference should tell readers which release changed which kind of work.

This watchlist was checked on 3 October 2026. It is a map of documented developments, not an independent ranking.

DeepSeek: follow the deployed model name

DeepSeek's current API documentation identifies deepseek-flash with V4.1 Flash and also lists V4 Pro. That makes exact endpoint and model-name checks important when reading an older comparison. A familiar alias does not necessarily mean the underlying model stayed unchanged.

My takeaway is practical: record the date and deployed version before comparing results. A comparison that silently spans two different versions is difficult to reproduce.

Qwen: image creation is its own story

Qwen-Image-2.1 was announced on 20 September. The release combines generation and editing and includes native transparent-image support. This belongs in a creative-workflow comparison, rather than a table that treats every new model as a text assistant.

The questions worth asking are about a usable asset: subject consistency, the edit requested, edge quality, and how much correction remains. A pretty sample is not enough to judge a production workflow.

Kimi: the product route matters

Kimi Code's September release notes describe K2.8 Preview and a desktop application. A reader should distinguish the model used by a service from the surrounding coding environment. Better task controls can improve a workflow even when they do not establish a new model-quality result.

The lesson is to compare both layers. What can the model do? What can the application let the user inspect, approve, resume, or undo? Those questions are related but different.

GLM: read the migration details

GLM-5.3's documentation lists text-only input and always-enabled reasoning with adjustable effort. That is a concrete implementation distinction. “New flagship” is less useful to a developer than knowing whether an existing request configuration remains valid.

For this release, I would examine the model's documented constraints before assuming it can replace another endpoint without changes. Similar-looking APIs can still expose different capabilities.

My view: avoid the country leaderboard

The origin of a model is relevant to its ecosystem, but it is a weak substitute for a task-specific comparison. Licensing, deployment options, data handling, availability, and output quality deserve their own columns. None can be inferred reliably from a provider's nationality.

The most useful coverage follows exact releases, keeps source dates visible, and reports what the announcement changes for readers. It also leaves room for uncertainty. These providers are worth following because they offer different approaches, not because one broad label makes them interchangeable.

Sources & further reading