Independent journal of China AIPrimary sources / editorial judgment
China AI DeskNumbered issues / primary sources

Daily desk

03 signals

What changed
what teams should do next.

This is a dated source check, not a claim that every underlying announcement was published today. Each brief separates the source date, the 22 July check, and the operational consequence.

01Build

Deadline · source dated 24 Apr · checked 22 Jul

Two days remain before DeepSeek’s legacy model IDs are scheduled to end.

DeepSeek’s official API change log says deepseek-chat and deepseek-reasoner will be discontinued on 24 July 2026. During the transition, those names point to the non-thinking and thinking modes of deepseek-v4-flash.

This is an application-contract change, not merely a naming cleanup. Teams should inspect hard-coded IDs, gateway aliases, fallback chains, monitoring labels, cost attribution, and tests that assume the two legacy names identify separate model families. A request returning successfully today does not prove that the application has migrated to the supported contract.

Do nowMove explicit model selection and reasoning-mode assumptions onto the supported V4 contract.
DeepSeek API change log Read the full Issue 03 analysis
02Model

Release state · official page checked 22 Jul

Kimi K3 is usable through the API; its full-weight state remains future-dated.

Moonshot’s official Kimi K3 page still says the full model weights will be released by 27 July 2026. The same page provides the kimi-k3 API call and says access unlocks after a minimum one-dollar top-up.

The two facts support different decisions. Product teams can test response behavior, tool use, latency, pricing, and integration now. Self-hosting teams still lack the released artifact set, final license terms, real memory profile, serving support, and independent reproduction needed for a deployment judgment. The five-day date is a checkpoint, not proof that an operational open-weight ecosystem already exists.

Do nowEvaluate the hosted model; wait for artifacts and terms before sizing a self-hosted deployment.
Kimi K3 official documentation
03Tools

Workflow · source dated 16 Jul · checked 22 Jul

Qwen Code is making learned skills and verified review part of the product layer.

Qwen Code’s 16 July update covers releases 0.19.9 through 0.19.11. It introduces a /learn command that turns a directory, URL, conversation, or text into a reusable skill; review suggestions can be posted as inline comments; and later review steps verify findings and reverse-audit the result.

The larger signal is not another isolated feature list. Chinese AI competition is moving from model quality alone toward the reliability of the surrounding workflow: reusable knowledge, review discipline, multi-workspace operation, artifacts, recovery, and human approval. For developer-tool teams, the relevant comparison is increasingly the full execution environment rather than the model endpoint hidden underneath it.

Test nextCompare review accuracy, learned-skill portability, and failure recovery—not only benchmark claims.
Qwen Code 16 July update Read the full Issue 04 analysis

Daily method

Short does not mean source-light.

The Daily Desk records a source date, a check date, the current reading, and the action it changes. Long-form interpretation remains in numbered issues.