Important updates from Kimi, the models built by Chinese lab Moonshot AI, taken from the company's own pages and rewritten in plain language. The log starts at the launch of Kimi K3. Sorted by when we found them, and every entry links to the source.
Kimi Code now accepts sign-in from both kimi.ai and kimi.com accounts. The release adds the WaitFor tool, which lets the agent wait for a background task to finish within the same turn instead of ending the turn and being started again. The official Kimi Datasource plugin gains 13 new sources, among them WHO, FAO, OECD, Eurostat and the European Central Bank. One rule tightens at the same time: Edit and Write now require the file to be read before it is changed.
Why it matters
WaitFor removes an artificial pause in long workflows, where the agent previously had to end its turn and be woken again to see the result of something it started itself. Requiring a read before a write guards against the agent overwriting content it has never seen.
Several skills can now be activated in a single request. A skill is a bundle of instructions for a particular kind of task, and you insert one by typing / after a space. The Windows single-binary build can now update itself. In the web interface the sidebar gains Open, Done and Workspaces tabs, and sessions can be marked as done.
Why it matters
Combining skills in one prompt means you no longer have to pick which one applies when a task spans several areas. Automatic updates on Windows remove a manual step that is easy to forget.
The setting for which model subagents use has grown into a pool. In config.toml you can list several models with a sentence describing what each is good for, and the main agent picks one based on the task. You can also pin the choice with force = true. The release adds an experimental fullscreen terminal mode and renders mathematical formulas written in LaTeX as ordinary characters.
Why it matters
A cheap, fast model is enough for small edits, while hard debugging needs the expensive one. With a pool you avoid picking one model for everything, and the descriptions you write are what the main agent actually reads when it decides.
The /tasks panel now shows the progress of subagents working in the background while they work. The release adds the Modern Web Guidance plugin to the bundled plugin catalog. It also fixes a bug where the token count reported after compaction fell far below the real size of the context.
Why it matters
Background agents that work with no visible progress are hard to trust, because you cannot tell whether they are working or stuck. The token bug is worth noting for anyone who was watching those numbers: they were wrong, not the context.
Following the Kimi K3 launch, kimi-k2.5 and the whole moonshot-v1 series are no longer available to newly registered users, and the platform shuts them down completely on August 31, 2026. That covers every moonshot-v1 variant, both text and vision. Moonshot AI recommends switching to a newer model as soon as possible. The documentation gives no date for when the notice itself was published.
Why it matters
This is a hard stop, not a recommendation: code still calling one of these models will fail after August 31. If you have a setup from before K3 arrived, it is worth checking which model it actually asks for.
Platform
Kimi Platform (the API)
First detected by us
August 22, 2026
Availability
Already closed to new registrations. Full shutdown on August 31, 2026.
Kimi Computer Use is the plugin that lets the model drive your computer, and it now runs on Windows x64 in addition to what it supported before. The release also adds a reminder that the cache is about to expire when you resume a session after a long idle, which can be turned off with the cache_expiry_hint setting. In the web interface, subagent tasks now show which model and thinking level they run with.
Why it matters
Windows users get access to the most far-reaching feature in Kimi Code. The cache reminder is practical: it explains why a resumed session suddenly costs more than it did.
Platform
Kimi Code CLI, version 0.34.0. The Computer Use plugin supports Windows x64.
First detected by us
August 7, 2026
Availability
Installed from /plugins. Subscription requirements are not specified.
Every command line surface in Kimi Code now runs on a new engine called agent-core-v2 by default, meaning the internal machinery driving a session. The old engine is still there behind an environment variable. The same release makes Kimi Computer Use and Kimi WebBridge official entries in the plugin marketplace, and /fork no longer switches you to the new branch: the original session keeps running, and the fork shows up in /sessions.
Why it matters
An engine swap is the kind of change that can produce new behavior without anything in the interface looking different, so it is worth knowing the fallback switch exists if something suddenly acts strangely. /fork leaving the original session alive means your background tasks do not stop when you branch.
Platform
Kimi Code CLI, version 0.33.0
First detected by us
August 7, 2026
Availability
Set KIMI_CODE_LEGACY_FLAG=1 to fall back to the old engine.
A hook is a small script that runs automatically when something specific happens. Kimi Code has gained four: TurnStarted, UserPromptQueued, TaskStarted and SessionHeartbeat, configured under [[hooks]] in config.toml. Two settings under [loop_control] have been renamed, from max_retries_per_step to max_attempts_per_step and from max_steps_per_run to max_steps_per_turn. The old names stop working but produce a warning at startup. A new [token_counting] section lets the tool estimate context usage itself when a provider does not report it.
Why it matters
The new hooks make it possible to attach your own automation at more points in a session. Note the rename: an existing setup does not carry on as before, it falls back to the defaults with a warning that is easy to miss.
Platform
Kimi Code CLI
First detected by us
August 5, 2026
Availability
Version 0.32.0. Subscription requirements are not specified.
Kimi Code is Moonshot's command line tool, where the model works directly in a code project from the terminal. Version 0.31.0 lets you define your own agents in ordinary markdown files, and plugins can now supply both ready-made agents and instructions for the system prompt. The /secondary_model command sets which model subagents use, though it is experimental and has to be enabled in /experiments first.
Why it matters
A team can package its own way of working into a plugin instead of every developer setting up the same thing again. The fix a day later removes, among other things, a bug where a new session displayed the wrong thinking level while actually running with thinking off.
Platform
Kimi Code CLI
First detected by us
July 31, 2026
Availability
Version 0.31.0, with fixes in 0.31.1 the following day. Subscription requirements are not specified.
Moonshot AI has published the complete weights for Kimi K3, meaning all the numbers that make up the finished trained model. With them, anyone can download the model and run it on their own hardware rather than going through the company's API. The date matches what the company promised in its launch post, which said the weights would arrive by July 27.
Why it matters
A model of this size that can be run locally is rare, and for organisations whose data cannot leave the building it is the difference between being able to use it and not. Actually running it demands serious hardware, and the licence is Moonshot's own rather than a standard open source one.
Platform
Hugging Face (moonshotai/Kimi-K3)
First detected by us
July 31, 2026
Availability
Open weights, uploaded July 27, 2026. Released under Moonshot's own Kimi K3 License, not a standard one.
Kimi K3 is Moonshot AI's new flagship model: 2.8 trillion parameters, a 1 million token context window, which is how much text it can hold in mind at once, and built-in understanding of images. The company calls it the world's first open model in the 3-trillion class. Moonshot states itself that overall it still trails Claude Fable 5 and GPT-5.6 Sol. The company's own pricing page lists kimi-k3 at $3 per million input tokens and $15 per million output tokens, with 30 cents for text already held in the cache.
Why it matters
The price sits well below what the leading American models charge, and the model can also be run locally by anyone who wants to. In exchange, the vendor itself concedes it is not as strong, so the choice is between cost and the last stretch of quality.
Platform
Kimi.com, Kimi Work, Kimi Code and the Kimi API
First detected by us
July 31, 2026
Availability
Available from launch day. The API model ID is kimi-k3.
This log starts on Jul 16, 2026. Anything older is deliberately left out, so it is not the vendor's full history.
