Kimi Release Notes

Minor Changes

• #2646 3c75a27 Thanks @liruifengv! - Show a cache-expiry reminder when resuming a long-idle session or submitting after a long idle stretch.

• #2697 e6e4ba2 Thanks @liruifengv! - web: When a model request fails and interrupts a conversation, a persistent failure card now stays in the session with one-click resume.

• #2652 68ba740 Thanks @wbxl2000! - Add Windows support for the built-in Kimi Computer Use capability and show the underlying error when capability setup fails. Install it from /plugins on Windows x64.

• #2697 e6e4ba2 Thanks @liruifengv! - web: Add a flat view to the sidebar session list.

Patch Changes …

Minor Changes

• #2407 0abcd00 Thanks @wbxl2000! - Add Kimi Computer Use and Kimi WebBridge as built-in official marketplace entries in the v2 CLI. Installing from /plugins sets up the latest managed runtime and plugin together, reports incomplete manual steps, and supports retrying interrupted setup.

• #2627 f881cdd Thanks @sailist! - Run the CLI surfaces (interactive TUI, kimi -p, kimi acp, kimi export, kimi provider) on the agent-core-v2 engine by default. Set KIMI_CODE_LEGACY_FLAG=1 to fall back to the legacy engine.

• #2565 54c04bf Thanks @7Sageer! - /fork no longer switches to the forked session: the current session stays active and its background tasks keep running. Find the fork in /sessions. …

Minor Changes

• #2558 75395f6 Thanks @sailist! - Add the TurnStarted, UserPromptQueued, TaskStarted, and SessionHeartbeat hook events, enrich hook payloads with the session title and client type, include the model and profile in SessionStart, and report SessionEnd as archive when a session is archived instead of exited. Configure the new events under [[hooks]] in config.toml.

Patch Changes

• #2416 eaab2b6 Thanks @mangeshraut712! - Fall back to the built-in models.dev catalog snapshot when the public catalog is unreachable, so Known third-party provider import still works offline or in blocked networks.

• #2083 bfa0080 Thanks @StaR4y! - web: Fix dark-mode monochrome controls and align the chat composer corner radius with the design system. …

Patch Changes

• #2410 f1a3475 Thanks @sailist! - Fix sporadic "model is not configured" errors when starting kimi web, caused by the background provider-model refresh transiently clearing the model catalog while the first session was being created.

• #2400 1f3f5da Thanks @7Sageer! - Preserve the assistant's partial output when a turn is interrupted with Esc, and remind the model that the previous turn was deliberately interrupted.

• #2415 5c0ec29 Thanks @wbxl2000! - web: Enable Monaco-based highlighting for code blocks, and fix line numbers overlapping or drifting out of alignment in fallback-rendered code blocks.

• #2442 bb2919e Thanks @liruifengv! - Reduce frequent full-screen redraws in the TUI. …

Minor Changes

• #2365 fa2c5ce Thanks @7Sageer! - Add support for plugin-contributed custom agents, discovered automatically and available for sub-agent delegation. Ship an agents/ directory in the plugin (or declare agents paths in the plugin manifest) to provide them.

• #2314 02d77b2 Thanks @7Sageer! - Allow enabled plugins to contribute agent system-prompt instructions through systemPrompt or systemPromptPath in kimi.plugin.json, effective on both agent engines (the TUI, kimi -p, and kimi web).

• #2232 efac96c Thanks @7Sageer! - Support Markdown-defined custom agents on agent-core.

• #2232 efac96c Thanks @7Sageer! - Add the /secondary_model slash command to configure the secondary model used by subagents.

Patch Changes …

Minor Changes

• #2255 67dd031 Thanks @he-yufeng! - Add a customizable footer status line, configured via [status_line] in tui.toml.

Patch Changes

• #2313 de0ba9d Thanks @starquakee! - Stop the turn after repeated invalid tool calls instead of retrying indefinitely.

• #2147 29783e4 Thanks @wbxl2000! - Show a quota note after installing official plugins that bill against plan quota (such as Kimi Datasource).

• #2147 29783e4 Thanks @wbxl2000! - Show a notice when an official plugin used in the session has an update available. Run /plugins to install it. …

New

• PPT slide editor is live: open and edit slides right in the workspace — changes take effect immediately
• Screenshot annotation is now supported for files in the preview area and in the browser; finished annotations can be sent directly to the Agent for revision
• Storage drive migration is now supported on Windows: Work data can be moved to another drive
• The Windows installer now supports choosing a custom install location

Changed

• Improved some token efficiency issues

Fixed

• Fixed several bugs and improved some interactions

Was this article helpful?

Patch Changes

• #2192 7799bd7 Thanks @sailist! - Hold per-agent runtime state of the experimental engine in the agent-scope state container, so it is observable in one place and disposed with the agent; state snapshots collapse class instances to name markers so resource graphs cannot exhaust memory during export.

• #2119 f06eb5c Thanks @pvzheroes125! - Allow hosts to defer registered user-tool schemas until needed. Set disclosure: "deferred" when registering a tool.

• #2192 7799bd7 Thanks @sailist! - Instantiate every registered service eagerly at scope creation on the experimental engine, following the dependency graph automatically, and drop the hand-maintained lists that resolved side-effect services one by one at startup. …

Patch Changes

• #2065 527d485 Thanks @7Sageer! - Add global default MCP server timeouts in config.toml and env vars.

• #2104 66f611a Thanks @RealKai42! - Fix loss of thinking content with OpenAI-compatible endpoints that return reasoning under a different field name (e.g. newer vLLM); the reasoning field is now detected per endpoint and echoed back on follow-up requests.

• #2089 ca38b7e Thanks @liruifengv! - Remove the toolbar tip that suggested trying the "superpowers" plugin.

• #2064 7b62ed5 Thanks @7Sageer! - Add experimental secondary-model bindings for newly spawned subagents, including per-agent model preferences and subagent-only model overrides. …

Minor Changes

• #1992 a8f1ca3 Thanks @RealKai42! - Support selecting a thinking effort level from ACP clients: the thinking picker now lists the current model's declared levels (for example off / low / medium / high) instead of only an on/off toggle. Use the thinking selector in your ACP client (e.g. Zed) to pick a level; the legacy on/off values keep working.

• #1735 ce0e3ce Thanks @7Sageer! - Let custom agent files restrict which sub-agent types they may delegate to (v2 engine only).

• #1735 ce0e3ce Thanks @7Sageer! - Support custom agents defined as Markdown files with frontmatter, usable as the main agent or a sub-agent (v2 engine only). …

Patch Changes

• #934 c5b6103 Thanks @tt-a1i! - Allow ACP sessions to start with configured non-OAuth model credentials instead of requiring terminal login.

• #1967 ad8cc85 Thanks @sailist! - Run web servers foreground-only end to end: the /web slash command now always starts a new server, and the kimi web kill / kimi web ps subcommands are removed — foreground servers stop with Ctrl+C. kimi server kill remains as a deprecated fallback that only stops servers started by a version before 0.28.0.

• #1948 f6f4192 Thanks @sailist! - Fix running subagents not observing permission mode switches made after they started.

Features ​

• Breaking:
  • The kimi server command tree is deprecated; use kimi web instead.
  • kimi web now runs in the foreground of the current terminal and opens the browser; stop it with Ctrl+C.

Polish ​

• Thinking effort persists only levels below the model's top tier (max).
• web: Add a note in the model switcher that switching models or thinking effort invalidates the existing prompt cache.

Bug Fixes ​

• Correct the YOLO and Auto permission mode descriptions: YOLO auto-approves tool actions but the agent may still ask questions, while Auto is fully autonomous and never asks.
• Fix the web backend ignoring symbolic links when loading AGENTS.md files and reading files.

Features ​

• Add the /copy slash command to copy the last assistant message to the clipboard.
• Using an API key for Kimi coding models now also fetches the latest model list automatically.

Polish ​

• OAuth connection errors now include the underlying network cause (DNS, refused connection, TLS, or timeout) instead of a bare "fetch failed".

Bug Fixes ​

• Fix repeated request rejections after an interrupted model response.
• Fix the built-in URL fetch tool's network safeguards: crafted domains and redirect chains can no longer reach loopback or internal network services.
• web: Fix LaTeX formulas rendering as garbled overlapping text when the web UI is accessed over the network.
• web: Fix queued messages silently re-sending previously uploaded files when a session is reopened.
• web: Remember the thinking level per model, fixing an empty, unresponsive thinking picker when the model doesn't support the stored level.
• web: Fix duplicate workspace groups on Windows when the same folder is opened with different path spellings; its sessions now list under one merged group.
• Fix AGENTS.md files installed as symbolic links being ignored by the web backend.
• Fix Esc and Ctrl+C cancelling compaction instead of closing an open /btw panel.
• Fix whitespace-only thinking content rendering as a blank line in the transcript. …

Polish ​

• Expand the coder subagent tool set to include background tasks, todo lists, plan mode, skill invocation, and nested agents, mirroring the main agent's capabilities.
• Warn in the /model and /effort pickers that switching invalidates the existing prompt cache, and hint to use /new to avoid extra token costs.
• web: Refresh the model catalog for all providers when opening the model picker, so newly available models always show up.
• Optimize the unit formatting of the context usage display.

Bug Fixes ​

• Fix a resumed session being marked as just updated and jumping to the top of the session list without any new activity.
• Fix the context size indicator under-reporting the model's actual context usage.
• Fix Kimi-provider models routed through the Anthropic protocol incorrectly showing reasoning effort options.
• Honor an explicit thinking "off" on OpenAI-compatible (chat completions) providers.
• Report when users stop tasks and preserve other stop reasons in model context.
• Fix a race where resuming a background subagent right after it was manually stopped could fail with an "already running" error.
• Replay empty thinking content verbatim instead of substituting a placeholder space on Anthropic-compatible and Kimi preserved-thinking endpoints.
• Keep legacy migrations idempotent across multiple Kimi homes and report damaged or unmapped sessions instead of silently skipping them. …

Features ​

• web: Attach any file type in chat — files can be dropped anywhere in the window, and sent files, images, and videos show as chips in the message bubble.

Polish ​

• web: Show full diagnostics for model request failures.
• Apply official Anthropic effort profiles and a 128k output fallback for unknown models.

Bug Fixes ​

• Fix the web server bearer-token check being bypassed by percent-encoded API paths, which allowed unauthenticated access to every API route.
• Fix the session filesystem API following symlinks that point outside the workspace, which allowed accessing host files beyond the session directory.
• web: Keep session activity indicators in sync with agent work and prevent duplicate streamed content after session activation races or LLM retries.
• Fix custom-named models on Anthropic-compatible providers starting new sessions with thinking effort off and not showing the thinking control in ACP clients.
• Honor adaptive_thinking = false on Anthropic-compatible models by omitting the effort parameter from requests.
• web: Fix the Content-Security-Policy on non-loopback server binds blocking the web UI's theme bootstrap script and bundled fonts.
• Fix sessions failing to be created when the workspace directory is given through a symlink.
• Fix the CLI exiting unexpectedly when reading an image from the clipboard fails; it now falls back to pasting text. …

Kimi K3, our most capable model to date, is now released and open-sourced, and available in Kimi Code.

Model scale & architecture ​

2.8 trillion parameters, built on KDA hybrid linear attention and Attention Residuals — the first open-source model at the 3-trillion-parameter scale.

Long context & multimodal ​

Native visual understanding with up to 1M-token context, handling large codebases and extended coherent reasoning.

Coding & agent strengths ​

Excels at long-horizon programming, game dev / 3D, and knowledge work; sustains long engineering tasks with minimal human supervision across kernel optimization, GPU compilers, chip design, and scientific computing.

Benchmark positioning ​

A top-tier model that outperforms Opus 4.8, GPT 5.5, and other mainstream models; on challenging coding tasks such as kernel optimization, it performs close to the strongest proprietary models.

Membership ​

Moderato members and above can use Kimi K3; Allegretto members and above unlock the 1M context window. K3 now supports low / high / max thinking effort levels, so you can pick the right intensity for each task. → Model configuration

Usage tip ​ …

Features ​

• Add a builtin /check-kimi-code-docs skill that automatically answers Kimi Code product questions with official-docs sources.

Polish ​

• Align kimi -p behavior across engines: print_background_mode and print_max_turns now apply, and /goal runs stay alive until the goal finishes.
• kimi -p now stays alive by default while background tasks are pending, with no effective wait or turn limit, and feeds each completion back to the agent. Set print_background_mode = "exit" or "drain" to restore the old exit-after-one-turn behavior.
• kimi -p background tasks and subagents no longer time out by default (interactive mode is unchanged); restore limits with [background] bash_task_timeout_s or [subagent] timeout_ms.
• Subagent timeout now defaults to 2 hours everywhere; override with [subagent] timeout_ms or KIMI_SUBAGENT_TIMEOUT_MS.
• The per-step LLM retry limit is raised from 3 to 10 attempts, so transient provider failures (429 / overload) are retried before a turn fails; tune with loop_control.max_retries_per_step.
• Workspaces now stay in sync: new sessions register automatically, missing workspaces are restored at startup, and removed ones stay removed.
• kimi web now logs failed requests and key operations so daemon issues are easier to diagnose.
• web: AgentSwarm cards now stay expanded while subagents are still running. …

Bug Fixes ​

• Fix Kimi sessions getting stuck when preserved-thinking history contains an empty reasoning step.
• Fix built-in tools being unavailable when the model provider becomes ready after the session starts.
• Fix Thinking effort routing: non-Kimi providers now preserve configured values, while Kimi models validate runtime selections and fall back safely during model resolution.
• web: Align thinking-level handling with the CLI: submit the selected level verbatim instead of silently downgrading it, fall back to the model's own default when nothing was chosen or the model switches, and persist explicit picks as the default for new sessions.
• Preserve goal completion summaries and show untyped LLM errors without an internal error-code prefix in step interruption events.

Polish ​

• web: Show just the level name (e.g. Max) in the model pill instead of "thinking: max".

Features ​

• web: Add session export: run /export or pick Export session from a session's more menu to download the session and troubleshooting logs as a ZIP (limited to 64 MiB).
• Move foreground Bash commands that hit their timeout to the background instead of killing them, so long-running commands survive the timeout and report back on completion. Set bash_auto_background_on_timeout = false under [background] in config.toml to restore the kill-on-timeout behavior.

Polish ​

• web: Refine goal mode controls with animated strip interactions, budget-aware progress, and design-system cancellation confirmation.
• On session close, background tasks are now asked to stop and given a grace period before being force-stopped.
• Rewrite repeated-tool-call reminders to redirect the agent toward a different action instead of prohibiting the call.
• Optimize the TaskOutput tool prompts to discourage blocking waits on background tasks.
• Send the kimi-code-cli User-Agent on provider registry (api.json) and model catalog fetches, so registries can identify the client version.
• Log a warning when a skill fails to parse instead of silently dropping it, and fix skill scan results not being reported.

Bug Fixes ​

• Prevent oversized image reads from poisoning sessions; sessions that already failed with request-too-large errors now recover automatically. …