Kimi Release Notes

Minor Changes

• #3434 ae7a6dc Thanks @sailist! - The kimi acp subcommand no longer honors KIMI_CODE_LEGACY_FLAG; it always runs on the default agent engine.
• #3334 971a8b2 Thanks @7Sageer! - The subagent model pool ([secondary_model]) is enabled by default in every launch mode and remains opt-out via KIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL=0 or [experimental] secondary-model = false.
• #3454 913a242 Thanks @wbxl2000! - web: Added a Plugins panel to Settings for browsing the plugin marketplace and installing, enabling, disabling, and removing plugins.

Patch Changes

• #3444 b4ae7f8 Thanks @sailist! - Remove the workspace restriction on the Bash tool's cwd parameter.
• #3454 913a242 Thanks @wbxl2000! - web: Added a code wrap toggle to the diff panel and streamlined its header.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed new content not appearing after auto-compaction (stuck at "Requesting…" above the divider); output after the compaction point now renders below the divider.
• #3392 616d510 Thanks @7Sageer! - Preserve comments, key order, and formatting in config.toml when configuration values are updated.
• #3348 9d2304c Thanks @liukx0205! - Fix models and providers transiently disappearing when config.toml is saved non-atomically by an external editor while the daemon reloads it.
• #3454 913a242 Thanks @wbxl2000! - web: The connecting splash screen now shows the current loading stage and the reason when the connection fails.
• #3290 4b9888b Thanks @sailist! - Block dangerous shell commands such as shutdown, reboot, or rm -rf in Auto mode, and always ask before running them in Manual and YOLO modes; disable the guard with [permission] dangerous_command_guard = false or KIMI_CODE_DANGEROUS_COMMAND_GUARD=false.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed queued or steered messages being unexpectedly retracted by pressing Esc after they start running.
• #3421 9c37feb Thanks @sailist! - Make session forks much faster.
• #3454 913a242 Thanks @wbxl2000! - web: Show a refresh button in the right-sidebar file preview when the previewed file is edited mid-turn.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed the server-side attachment notice text leaking into your own message bubble after sending a file.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed misaligned content in the composer attachment tooltip.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed messages sent with Ctrl+S while the agent is running disappearing after a page reload.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed sessions occasionally stuck showing "working" after a message is appended mid-turn.
• #3377 58b74cf Thanks @chengluyu! - Fix duplicate user messages in transcript clients.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed the first pinyin letter being committed as English text in Chinese IMEs after enabling goal or plan mode.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed sessions occasionally stuck showing "working" long after the reply had completed.
• #3427 442b563 Thanks @sailist! - Honor explicit [experimental] config entries over the KIMI_CODE_EXPERIMENTAL_FLAG master switch, so a flag set to false in config.toml stays off; per-feature KIMI_CODE_EXPERIMENTAL_<NAME> variables still override both.
• #3412 7bc5b20 Thanks @7Sageer! - Send the forked-subagent context notice as a system reminder.
• #3415 82bf0a8 Thanks @kimi-agent-bot! - Parse git status --porcelain with -z so non-ASCII paths are no longer mangled into bogus quoted directory segments.
• #3454 913a242 Thanks @wbxl2000! - web: Pressing Tab in the @ file menu completes the highlighted candidate's name into the input while keeping the menu open for further filtering.
• #3454 913a242 Thanks @wbxl2000! - web: Support activating multiple skills from a single message.
• #3454 913a242 Thanks @wbxl2000! - web: Background task notification cards now show a "Sent from background · bash" source line and a single "status: task description" body line.
• #3454 913a242 Thanks @wbxl2000! - web: Background task notification cards now label the sender as "Sent from background (Bash) / (Agent)".
• #3454 913a242 Thanks @wbxl2000! - web: Superseded @ file searches are now cancelled promptly during fast typing, reducing background load.
• #3371 9e88152 Thanks @tpoisonooo! - Fix prompts remaining queued forever after reopening a session.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed manually typed > quote blocks in messages being misrendered as quote annotations; they now render as plain text.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed messages rejected by the server sometimes showing no failure toast.
• #3425 ceb5153 Thanks @sailist! - Add the kimi session list command to list sessions from the command line.
• #3390 76c1a7a Thanks @Grapedge! - Simplify the built-in system prompt.
• #3454 913a242 Thanks @wbxl2000! - web: Selecting a skill from the slash menu now inserts the same skill pill as the @ menu.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed stale steered-message bubbles lingering after a transcript refresh.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed the file-change summary card appearing too early when a turn is steered while still running.
• #3436 0f39b2c Thanks @wbxl2000! - Fix idle sessions briefly showing a "Working" state when opened in desktop and web clients.
• #3346 ece9618 Thanks @tpoisonooo! - Tower mode (experimental, KIMI_CODE_EXPERIMENTAL_TOWER=1): spawned workers now start from the base checkout's uncommitted changes instead of missing them, and TowerMerge refuses to merge while the checkout still holds those changes uncommitted. Also, a new session can now enter tower mode after the previous owning session stopped without exiting, instead of being refused while that session stays open. Tower mode now stays on after tower teardown; turn it off explicitly with /tower off. Tower mode is now mutually exclusive with plan mode and swarm mode: entering any one of them exits the others.
• #3399 c3bf6f9 Thanks @tpoisonooo! - Tower mode (experimental, KIMI_CODE_EXPERIMENTAL_TOWER=1): the agent can no longer enter tower mode on its own — turn it on with /tower on, or with /tower (also in the web UI) to pin the local branch missions merge back into; a missing base branch is created from the current checkout (uncommitted changes committed onto it as a labeled WIP snapshot) and the workspace is initialized or rebased to it immediately, refusing with guidance while missions are open. Tower agents that die (failed, timed out, killed, or lost) are recorded in the tower protocol — TowerStatus marks them in the roster and warns about missions whose owner died, with a resume hint — and the tower's console instructions now require summarizing every worker's deliverables per mission before teardown.
• #3454 913a242 Thanks @wbxl2000! - web: Fixed the context usage ring not refreshing after compaction; session usage now updates live from transcript metadata.
• #3391 5f0aa7f Thanks @7Sageer! - Default the workspace trust prompt selection to "Trust this folder" instead of "Don't trust".
• #3366 9619277 Thanks @kimi-agent-bot! - Use the Unicode ellipsis "…" in user-facing TUI and VS Code webview text.
• #3405 630a11d Thanks @kimi-agent-bot! - Download compressed native update artifacts and decompress them while staging.

Original source

Features

• web: Add a Plugins panel to Settings for browsing the plugin marketplace and installing, enabling, disabling, and removing plugins.
• web: Support activating multiple skills from a single message.
• Add the kimi session list command to list sessions from the command line.
• Tower mode (experimental, KIMI_CODE_EXPERIMENTAL_TOWER=1): the agent no longer enters tower mode on its own — turn it on with /tower on or /tower .
• The subagent model setting ([secondary_model]) graduates from experimental to stable.
• Block dangerous shell commands such as shutdown, reboot, or rm -rf in Auto mode, and always ask before running them in Manual and YOLO modes; disable the guard with [permission] dangerous_command_guard = false or KIMI_CODE_DANGEROUS_COMMAND_GUARD=false.

Polish

• Preserve comments, key order, and formatting in config.toml when configuration values are updated.
• Remove the workspace restriction on the Bash tool's cwd parameter.
• Default the workspace trust prompt selection to "Trust this folder" instead of "Don't trust".
• The kimi acp subcommand no longer honors KIMI_CODE_LEGACY_FLAG; it always runs on the default agent engine.
• web: Add a code wrap toggle to the diff panel and streamline its header.

Bug Fixes

• Honor explicit [experimental] config entries over the KIMI_CODE_EXPERIMENTAL_FLAG master switch, so a flag set to false in config.toml stays off; per-feature KIMI_CODE_EXPERIMENTAL_ variables still override both.
• Fix several known issues and make various refinements. See the changelog on GitHub for more technical entries.

Original source

Renamed permission modes and a new dangerous-command guard

The former YOLO / Auto modes are now Ask When Needed (/ask-when-needed) and Never Ask (/never-ask); Manual is Always Ask. A built-in dangerous-command guard also ships in this release: commands such as shutdown, reboot, or rm -rf are blocked outright in Never Ask mode and always require your confirmation in the other modes; disable it with [permission] dangerous_command_guard = false. → Interaction & input

Subagent model pool graduates from experimental

The [secondary_model] subagent settings are now generally available and enabled by default in every launch mode; you can still opt out with KIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL=0 or [experimental] secondary-model = false. → Config files

Plugins panel in web settings

The web settings now include a Plugins panel for browsing the plugin marketplace and installing, enabling, disabling, or removing plugins. → Plugins

Also in this release: Tower mode (experimental) no longer starts on its own — turn it on explicitly with /tower on or /tower ; the new kimi session list command lists sessions from the command line. v0.40.1 fixes the kimi-cli migration prompt reappearing after the migration had completed or been dismissed.

Original source

Patch Changes

• #3333 8f43674 Thanks @kimi-agent-bot! - web: Fix the command tool row rendering noticeably taller than other tool rows.
• #3333 8f43674 Thanks @kimi-agent-bot! - web: Fix the first IME (or keyboard) character being silently swallowed after clicking the placeholder text in an empty composer.
• #3333 8f43674 Thanks @kimi-agent-bot! - web: Fix switching the permission mode in one session changing it for every session; the permission mode is now scoped per session.
• #3333 8f43674 Thanks @kimi-agent-bot! - web: Fix flickering and broken interactions in the image and video attachment preview popovers.
• #3333 8f43674 Thanks @kimi-agent-bot! - web: Fix attachments in a newly created session still showing as uploading after the upload has finished.
• #3333 8f43674 Thanks @kimi-agent-bot! - web: Render the rich composer placeholder outside the editor, fixing the first typed/IME character being swallowed.
• #3307 0310f22 Thanks @kimi-agent-bot! - Increase the request timeout for kimi update.
• #3333 8f43674 Thanks @kimi-agent-bot! - web: Unify right-side panel headers and give the OpenIn menu a file mode (copy absolute path, editor picker, full-path tooltip).
• #3333 8f43674 Thanks @kimi-agent-bot! - web: Fix signed-in users without a usable model being wrongly asked to sign in (and getting stuck there on web); the send gate now offers picking or configuring a model instead.
• #3333 8f43674 Thanks @kimi-agent-bot! - web: Fix startup getting stuck on "Connecting…" for a long time when the account has many workspaces.
• #3328 dc6028d Thanks @kimi-agent-bot! - Fixed skill instructions injected by the Skill tool showing up as ordinary user messages in the rebuilt transcript.
• #3292 23921e9 Thanks @7Sageer! - When a session resumes, the assistant is warned that background tasks from the previous session may still be running.

Original source

Bug Fixes

• web: Fix switching the permission mode in one session changing it for every session; the permission mode is now scoped per session.
• web: Fix signed-in users without a usable model being wrongly asked to sign in (and getting stuck there on web); the send gate now offers picking or configuring a model instead.
• web: Fix the first IME (or keyboard) character being silently swallowed after clicking the composer placeholder.
• web: Fix attachments in a newly created session still showing as uploading after the upload has finished.
• Fix several known issues and make various refinements. See the changelog on GitHub for more technical entries.

Original source

Features

• Add Remote Control as an experimental feature for accessing a local web session remotely. Enable it with KIMI_CODE_EXPERIMENTAL_REMOTE_CONTROL=1, then run kimi rc, kimi web --remote-control, or /remote-control to start it.
• Add experimental tower mode for multi-agent orchestration; set KIMI_CODE_EXPERIMENTAL_TOWER=1, then run /tower on and /tower to start.
• Add an optional fork parameter to subagent and swarm tools that starts the subagent with a snapshot of the calling agent's conversation history; set KIMI_CODE_EXPERIMENTAL_SUBAGENT_FORK=1 or subagent_fork = true under [experimental] in config.toml to enable it.
• web: Allow moving a running foreground Bash command or subagent to the background via the "Move to background" button on the running card.
• web: Add a flat/by-workspace tab to the mobile session list.
• Add the Tencent CloudBase plugin to the curated marketplace.
• Add a dedicated [swarm] timeout_ms config option (or the KIMI_CODE_SWARM_TIMEOUT_MS env var) for AgentSwarm subagent timeouts, which no longer follow [subagent] timeout_ms.

Polish

• web: Revamp the right sidebar as a multi-tab panel.
• web: Improve composer interaction, including the presentation of file, folder, and media attachments.
• web: Improve mobile UI styling.

Bug Fixes

• Fix file tools and shell working directories failing to resolve Git Bash paths such as /c/Users or /tmp on Windows.
• Fix several known issues and make various refinements. See the changelog on GitHub for more technical entries.

Original source

Remote control for local web sessions

Set KIMI_CODE_EXPERIMENTAL_REMOTE_CONTROL=1, then run kimi rc, kimi web --remote-control, or /remote-control to access a local web session remotely.

Experimental Tower multi-agent orchestration

Set KIMI_CODE_EXPERIMENTAL_TOWER=1, then use /tower on and /tower to start Tower mode and coordinate multiple agents toward an objective.

Fork subagents from the current conversation

Set KIMI_CODE_EXPERIMENTAL_SUBAGENT_FORK=1, or subagent_fork = true under [experimental], to let the subagent and swarm tools use fork to start from a snapshot of the calling agent's conversation history. → Agents and subagents

Also in this release: Configure an independent AgentSwarm timeout with [swarm] timeout_ms or KIMI_CODE_SWARM_TIMEOUT_MS; file tools and Shell now resolve Git Bash paths correctly on Windows. v0.39.1 fixes the web permission mode leaking across sessions, the first typed character being swallowed in an empty composer, and startup hanging on "Connecting…" with many workspaces, and adds a file mode to the OpenIn menu.

Original source

Minor Changes

• #3034 f0a6094 Thanks @sailist! - Add Remote Control as an experimental feature for accessing a local web session remotely. Enable it with KIMI_CODE_EXPERIMENTAL_REMOTE_CONTROL=1, then run kimi rc, kimi web --remote-control, or /remote-control to start it.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix the slash-command and @-mention panels failing to open on mobile — both panels and the + menu are now grab-handle bottom sheets on small screens.
• #3157 491ebd0 Thanks @liruifengv! - web: Add a flat/by-workspace tab to the mobile session list.
• #3296 df9e858 Thanks @liruifengv! - web: Revamp the right sidebar as a multi-tab panel.
• #3007 f6736d7 Thanks @7Sageer! - Add an optional fork parameter to subagent and swarm tools that starts the subagent with a snapshot of the calling agent's conversation history; set KIMI_CODE_EXPERIMENTAL_SUBAGENT_FORK=1 or subagent_fork = true under [experimental] in config.toml to enable it.
• #3296 df9e858 Thanks @liruifengv! - web: Allow moving a running foreground Bash command or subagent to the background via the "Move to background" button on the running card.
• #3099 0f44537 Thanks @wbxl2000! - Add experimental tower mode for multi-agent orchestration; set KIMI_CODE_EXPERIMENTAL_TOWER=1, then run /tower on and /tower <objective> to start.

Patch Changes

• #3241 1dc34b4 Thanks @tpoisonooo! - Silence the MaxListenersExceededWarning that could appear during long agent turns with many parallel tool calls.
• #3166 d4e0ad4 Thanks @liruifengv! - web: fix thinking blocks in the subagent detail panel being stuck expanded and not collapsible.
• #3139 381142a Thanks @sailist! - Fix sessions failing to archive when their workspace folder no longer exists.
• #3034 f0a6094 Thanks @sailist! - Fix messages sent from one web client not appearing on other clients connected to the same session.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix pressing Esc to cancel an IME candidate also closing the BTW side chat.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix the composer not receiving focus after opening the BTW side chat via the shortcut or /btw.
• #3212 a664226 Thanks @sailist! - Preserve the active session and its selected model when logging out of a provider.
• #3136 e9a99e5 Thanks @wbxl2000! - Add the Tencent CloudBase plugin to the curated marketplace.
• #3296 df9e858 Thanks @liruifengv! - web: Improve code block interaction and rendering.
• #3296 df9e858 Thanks @liruifengv! - web: Improve composer interaction, including the presentation of file, folder, and media attachments.
• #3152 3090c1c Thanks @liruifengv! - web: Fix composer toolbar buttons squeezing and overlapping each other in very narrow windows.
• #3154 54bb49e Thanks @Grapedge! - Fix the latest reply disappearing from the transcript after a scheduled cron reminder fires.
• #3034 f0a6094 Thanks @sailist! - Remove the --allow-remote-terminals flag from kimi web; PTY terminal routes now stay available on loopback binds only.
• #3183 2adc6a1 Thanks @sailist! - Fix ACP session regressions: Bash, Grep, and Glob failing when the editor does not support terminal command execution, session creation failing with stdio MCP servers, and reopening a closed session failing with an internal error.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix memory usage growing steadily after repeatedly switching sessions and toggling the side chat and subagent panels.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix unsent composer attachments such as images being lost after switching sessions on the new-session page.
• #3296 df9e858 Thanks @liruifengv! - Fix known issues.
• #3294 21f7ef6 Thanks @liruifengv! - Fix sign-in briefly showing a device-code-expired error after a successful authorization.
• #3152 3090c1c Thanks @liruifengv! - web: Fix long question text in question cards being truncated with an ellipsis instead of wrapping.
• #3206 4d5147b Thanks @sailist! - Fix repeated server crashes when resuming a session that was interrupted in the middle of a turn.
• #3159 ea0626a Thanks @pvzheroes125! - Prevent AskUserQuestion from starting background tasks when task controls are unavailable.
• #3157 491ebd0 Thanks @liruifengv! - web: Remove the redundant parenthesized domain from the login entry card titles.
• #3234 74d9bd1 Thanks @xpzouying! - Send MCP structuredContent to the model only when the tool result has no usable content, avoiding duplicate tool output.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix the send and stop button icons rendering too small in the mobile composer.
• #3157 491ebd0 Thanks @liruifengv! - web: Present the mobile model picker as a bottom sheet consistent with the other mobile drawers.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix the oversized appearance theme cards in the mobile first-run wizard.
• #3157 491ebd0 Thanks @liruifengv! - web: Temporarily remove the custom-provider entry from the mobile first-run wizard.
• #3152 3090c1c Thanks @liruifengv! - web: Improve mobile UI styling.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix tool-call rows alternating heights on mobile by unifying them to the compact row height.
• #3152 3090c1c Thanks @liruifengv! - web: Collapse the composer model picker to an icon when space is tight; hovering still shows the model and reasoning effort.
• #3152 3090c1c Thanks @liruifengv! - web: Fix the composer permission mode label being hidden even when there is enough space.
• #3219 d1a46db Thanks @kimi-agent-bot! - Show the /plugins marketplace catalog as soon as it loads, with latest-version lookups running in the background.
• #3002 d723cc4 Thanks @7Sageer! - Respect workspace trust and configuration readiness when managing MCP servers.
• #3191 ee53d84 Thanks @7Sageer! - Fix subagents bound to a configured secondary model ignoring its default thinking effort.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix mismatched left and right margins in the sidebar session list, and show the scrollbar only while hovering or scrolling.
• #3164 41a75ad Thanks @kimi-agent-bot! - Fix the context usage bar in /usage and the footer showing a stale percentage after the context size or model changes.
• #3198 496bb6c Thanks @sailist! - Add a dedicated [swarm] timeout_ms config option (or the KIMI_CODE_SWARM_TIMEOUT_MS env var) for AgentSwarm subagent timeouts, which no longer follow [subagent] timeout_ms.
• #3152 3090c1c Thanks @liruifengv! - web: Restyle background task notifications as a lighter notice that shows the task summary, output files, and output preview directly.
• #3239 6595955 Thanks @wbxl2000! - Fix foreground subagents being reported as background tasks on the task list.
• #3227 4b04492 Thanks @7Sageer! - Save oversized tool output within safety limits for later inspection, report omitted MCP content, and retain partial assistant responses when streams fail.
• #3102 2f12469 Thanks @wbxl2000! - Fix the cold transcript rebuild splitting a turn at background-task completion notices; they now fold into the current turn like the live stream does.
• #3271 75c94c4 Thanks @sailist! - Fix attached images disappearing from the user message while the agent is working.
• #3278 b17bd61 Thanks @kimi-agent-bot! - Fix the "manually stopped" state lingering after undoing the interrupted turn.
• #3157 491ebd0 Thanks @liruifengv! - web: Fix the reset-time hint in the sidebar usage panel being ellipsized even when there is enough room.
• #2200 30e7f62 Thanks @wszqkzqk! - Fix file tools and shell working directories failing to resolve Git Bash paths such as /c/Users or /tmp on Windows.
• #3281 7de7b18 Thanks @sailist! - Fix sessions failing to resume when their session journal was truncated or corrupted, for example after a full disk.

Original source

Minor Changes

• #2862 3d77620 Thanks @liruifengv! - Support two OAuth login methods — kimi.ai and kimi.com.
• #3060 8440801 Thanks @chengluyu! - Add the WaitFor tool: the agent can now wait for a background task to finish within the current turn instead of ending the turn and being re-invoked.

Patch Changes

• #3135 2c5415f Thanks @liruifengv! - web: Label inline subagent cards in the message stream with their foreground or background mode.
• #3121 3899079 Thanks @7Sageer! - Fix config.toml entries being lost when the file had a syntax error or was edited outside the app.
• #3135 2c5415f Thanks @liruifengv! - web: Add copy buttons next to the server version and server address in settings.
• #3119 a34d02a Thanks @wbxl2000! - Add 13 data sources to the official Kimi Datasource plugin — Chinese government data (NDA/NBS) and standards (GB/HB/DB/TT), eight international organization datasets (WHO, FAO, UNSD, ECB, Eurostat, UNICEF, OECD, FRED), Xinhua Finance, and Caixin. Update the plugin from the Official tab in /plugins.
• #3096 67fbcdf Thanks @sailist! - Edit and Write now require reading an existing file before modifying it, and reject the write when the file changed on disk since it was last read.
• #3101 d96b4a0 Thanks @kimi-agent-bot! - Stop retrying requests blocked by the provider content filter; the filter notice now shows immediately.
• #3135 2c5415f Thanks @liruifengv! - web: Keep empty workspace groups visible in the legacy sidebar after their last session is archived.
• #3135 2c5415f Thanks @liruifengv! - web: Hide button hover tooltips outside a menu while the menu is open.
• #3135 2c5415f Thanks @liruifengv! - web: Keep the model picker menu on the workspace home within the viewport.
• #3135 2c5415f Thanks @liruifengv! - web: Fix the workspace group title showing untranslated text in the search dialog.
• #3135 2c5415f Thanks @liruifengv! - web: Fix the settings dialog dropdown list being clipped by the scroll area, and lock the content behind it while the list is open.
• #3135 2c5415f Thanks @liruifengv! - web: Keep the slash command and @ mention panels on the workspace home within the viewport.
• #3052 6595a69 Thanks @RealKai42! - Fix 422 errors from some OpenAI-compatible providers when a conversation includes tool calls.
• #3135 2c5415f Thanks @liruifengv! - web: Prevent text selection in the sidebar user menu and its plan usage submenu.
• #3135 2c5415f Thanks @liruifengv! - web: Fix slow session list loading when there are many workspaces.
• #3135 2c5415f Thanks @liruifengv! - web: Auto-open the browser authorization page after choosing a login region, redesign the authorization waiting page, and refresh the login state as soon as the window regains focus instead of waiting for the poll.
• #3083 571bcc2 Thanks @7Sageer! - Fix the missing OAuth authenticate tool for remote MCP servers that require login.
• #3135 2c5415f Thanks @liruifengv! - web: Upgrade the @ mention menu: file and skill candidates are merged and ranked by match quality, file search is faster, with path-fragment matching and hit highlighting.
• #3135 2c5415f Thanks @liruifengv! - web: Round menu items concentric with their menu frames.
• #3135 2c5415f Thanks @liruifengv! - web: Add a Pin action to the chat header more-menu to pin the current session to the sidebar pinned section.
• #3135 2c5415f Thanks @liruifengv! - web: Allow dragging the divider between the pinned section and the session list to resize both areas, with fade hints at the edges when the pinned section scrolls.
• #3135 2c5415f Thanks @liruifengv! - web: Improve the prompt queue interaction, with per-row steer and send.
• #3135 2c5415f Thanks @liruifengv! - web: Add kimi.com and kimi.ai OAuth login entries, and switch update and help links to the site matching the current login.
• #3135 2c5415f Thanks @liruifengv! - web: Remove sessions archived from another client from the session list immediately, without a manual refresh.
• #3135 2c5415f Thanks @liruifengv! - web: Label the timestamp at the bottom of the session menu as last active and tighten that row's padding.
• #3054 cfc3350 Thanks @Grapedge! - Collapse long ! shell command output instead of flooding the transcript. Press ctrl+o to expand or collapse it together with tool output.
• #3135 2c5415f Thanks @liruifengv! - web: Fix misaligned action buttons between the sidebar section headers and the session rows.
• #3135 2c5415f Thanks @liruifengv! - web: Remove the skill-activated card from skill activation messages.
• #3135 2c5415f Thanks @liruifengv! - web: Make skill-activation turns undoable so they can be withdrawn and resent.
• #3012 ca87c58 Thanks @sailist! - Sub-agents no longer spawn their own sub-agents by default; custom agent profiles can still allow it explicitly.
• #3005 be8e017 Thanks @wbxl2000! - Fix background agent rows that could not be stopped right after they appeared, and stray rows left behind when an agent failed to start.
• #3046 f13f379 Thanks @7Sageer! - Fix the model being directed to unavailable tools when it encounters an image or binary file.
• #3108 05f2ad5 Thanks @kimi-agent-bot! - web: clearing a goal now removes it from the transcript view instead of leaving the stale goal displayed.
• #3108 05f2ad5 Thanks @kimi-agent-bot! - web: attachments sent with a prompt now appear in the live transcript immediately instead of only after a reload.
• #3135 2c5415f Thanks @liruifengv! - web: Tighten the row height and spacing of the account menu and its submenus to match the standard menu density.
• #3135 2c5415f Thanks @liruifengv! - web: Give WaitFor tool calls a dedicated quiet-line display showing completed tasks, wait timeouts, and how many tasks are still running.

Original source

Features

• Support two OAuth login methods — kimi.ai and kimi.com.
• Add the WaitFor tool: the agent can now wait for a background task to finish within the current turn instead of ending the turn and being re-invoked.
• Add 13 data sources to the official Kimi Datasource plugin — Chinese government data (NDA/NBS) and standards (GB/HB/DB/TT), eight international organization datasets (WHO, FAO, UNSD, ECB, Eurostat, UNICEF, OECD, FRED), Xinhua Finance, and Caixin. Update the plugin from the Official tab in /plugins.
• web: Add a Pin action to the chat header more-menu.

Polish

• Edit and Write now require reading an existing file before modifying it.
• Collapse long ! shell command output instead of flooding the transcript. Press ctrl+o to expand or collapse it together with tool output.

Bug Fixes

• Fix config.toml entries being lost when the file had a syntax error or was edited outside the app.
• Fix several known issues and make various refinements. See the changelog on GitHub for more technical entries.

Original source

New WaitFor tool

The agent can now wait for a background task to finish within the current turn — no need to end the turn and get woken up again. → Background Tasks

13 new data sources in Kimi Datasource

The official data plugin adds Chinese government data (NDA/NBS) and standards (GB/HB/DB/TT), eight international organization datasets (WHO, FAO, UNSD, ECB, Eurostat, UNICEF, OECD, FRED), Xinhua Finance, and Caixin. Update the plugin from the Official tab in /plugins . → Kimi Datasource

Also in this release

Edit and Write now require reading an existing file before modifying it; overly long ! shell command output is collapsed automatically — press ctrl+o to expand it together with tool output.

Original source

Today, we are introducing Kimi K2.5, the most powerful open-source model to date.

Kimi K2.5 builds on Kimi K2 with continued pretraining over approximately 15T mixed visual and text tokens. Built as a native multimodal model, K2.5 delivers state-of-the-art coding and vision capabilities and a self-directed agent swarm paradigm.

For complex tasks, Kimi K2.5 can self-direct an agent swarm with up to 100 sub-agents, executing parallel workflows across up to 1,500 tool calls. Compared with a single-agent setup, this reduces execution time by up to 4.5x. The agent swarm is automatically created and orchestrated by Kimi K2.5 without any predefined subagents or workflow.

Kimi K2.5 is available via Kimi.com, the Kimi App, the API, and Kimi Code. Kimi.com & Kimi App now supports 4 modes: K2.5 Instant, K2.5 Thinking, K2.5 Agent, and K2.5 Agent Swarm (Beta). Agent Swarm is currently in beta on Kimi.com, with free credits available for high-tier paid users.

Across three agentic benchmarks—HLE, BrowseComp, and SWE-Verified—Kimi K2.5 delivers strong performance at a fraction of the cost.

1. Coding with Vision

Kimi K2.5 is the strongest open-source model to date for coding, with particularly strong capabilities in front-end development.

K2.5 can turn simple conversations into complete front-end interfaces, implementing interactive layouts and rich animations such as scroll-triggered effects. Beyond text prompts, K2.5 excels at coding with vision. By reasoning over images and video, K2.5 improves image/video-to-code generation and visual debugging, lowering the barrier for users to express intent visually.

K2.5 excels in real-world software engineering tasks. We evaluate it using Kimi Code Bench, our internal coding benchmark covering diverse end-to-end tasks — from building to debugging, refactoring, testing, and scripting — across multiple programming languages. On this benchmark, K2.5 shows consistent and meaningful improvements over K2 across task types.

To try out K2.5's agentic coding capabilities, K2.5 Agent offers a set of preconfigured tools for immediate, hands-on experiences. For software engineering use cases, we recommend pairing Kimi K2.5 with our new coding product, Kimi Code. Kimi Code works in your terminal and can be integrated with various IDEs including VSCode, Cursor, Zed, etc. Kimi Code is open-sourced and supports images and videos as inputs. It also automatically discovers and migrates existing skills and MCPs into your working environment in Kimi Code.

Here's an example using Kimi Code to translate the aesthetic of Matisse's La Danse into the Kimi App. This demo highlights a breakthrough in autonomous visual debugging. Using visual inputs and documentation lookup, K2.5 visually inspects its own output and iterates on it autonomously. It creates an art-inspired webpage created end to end.

2. Agent Swarm

Scaling Out, Not Just Up.

We release K2.5 Agent Swarm as a research preview, marking a shift from single-agent scaling to self-directed, coordinated swarm-like execution.

Trained with Parallel-Agent Reinforcement Learning (PARL), K2.5 learns to self-direct an agent swarm of up to 100 sub-agents, executing parallel workflows across up to 1,500 coordinated steps, without predefined roles or hand-crafted workflows.

PARL uses a trainable orchestrator agent to decompose tasks into parallelizable subtasks, each executed by dynamically instantiated, frozen subagents. Running these subtasks concurrently significantly reduces end-to-end latency compared to sequential agent execution.

Training a reliable parallel orchestrator is challenging due to delayed, sparse, and non-stationary feedback from independently running subagents. A common failure mode is serial collapse, where the orchestrator defaults to single-agent execution despite having parallel capacity. To address this, PARL employs staged reward shaping that encourages parallelism early in training and gradually shifts focus toward task success.

The performance reward evaluates the overall success and quality of the solution for a given task. This is augmented by two auxiliary rewards, each addressing a distinct challenge in learning parallel orchestration. The reward for parallel instantiation mitigates serial collapse by incentivizing subagent instantiation and exploration of concurrent scheduling spaces. The finish reward focuses on successful completion of assigned subtasks to prevent spurious parallelism, where the orchestrator spawns many subagents without meaningful task decomposition. Hyperparameters controlling these rewards are annealed to zero over training.

To further force parallel strategies to emerge, a computational bottleneck makes sequential execution impractical. Instead of counting total steps, performance is evaluated using Critical Steps, a latency-oriented metric inspired by the critical path in parallel computation. This metric captures orchestration overhead and reflects the slowest subagent at each stage. Under this metric, spawning more subtasks only helps if it shortens the critical path.

K2.5 Agent Swarm improves performance on complex tasks through parallel, specialized execution. In internal evaluations, it leads to an 80% reduction in end-to-end runtime while enabling more complex, long-horizon workloads.

Agent Swarm reduces the minimum critical steps required to achieve target performance by 3×–4.5× compared to single-agent execution in wide search scenarios, with savings scaling as targets rise—translating to up to 4.5× wall-clock time reduction via parallelization.

Representative trajectories demonstrate K2.5 Agent Swarm in action: Parallel at Scale with 100 Sub-agents Hunting for Creators. The task is to identify the top three YouTube creators across 100 niche domains. K2.5 Agent Swarm first researches and defines each domain, then autonomously creates 100 sub-agents to conduct parallel searches. Each sub-agent identifies leading creators within its assigned niche, and the results—300 YouTuber profiles—are aggregated into a structured spreadsheet.

3. Office Productivity

Kimi K2.5 brings agentic intelligence into real-world knowledge work.

K2.5 Agent can handle high-density, large-scale office work end to end. It reasons over large, high-density inputs, coordinates multi-step tool use, and delivers expert-level outputs: documents, spreadsheets, PDFs, and slide decks—directly through conversation.

With a focus on real-world professional tasks, two internal expert productivity benchmarks were designed. The AI Office Benchmark evaluates end-to-end Office output quality, while the General Agent Benchmark measures multi-step, production-grade workflows against human expert performance. Across both benchmarks, K2.5 shows 59.3% and 24.3% improvements over K2 Thinking, reflecting stronger end-to-end performance on real-world tasks.

K2.5 agent supports advanced tasks such as adding annotations in Word, constructing financial models with Pivot Tables, and writing LaTeX equations in PDFs, while scaling to long-form outputs like 10,000-word papers or 100-page documents. Tasks that once took hours or days now complete in minutes.

4. Conclusion

Grounded in advances in coding with vision, agent swarms, and office productivity, Kimi K2.5 represents a meaningful step toward AGI for the open-source community, demonstrating strong capability on real-world tasks under real-world constraints. Looking ahead, we will push further into the frontier of agentic intelligence, redefining the boundaries of AI in knowledge work.

Appendix

Benchmark table

To reproduce official Kimi-K2.5 benchmark results, we recommend using the official API. For third-party providers, refer to Kimi Vendor Verifier (KVV) to choose high-accuracy services.

Footnotes

1. General Testing Details

• Results reported for Kimi K2.5 and DeepSeek-V3.2 with thinking mode enabled, Claude Opus 4.5 with extended thinking mode, GPT-5.2 with xhigh reasoning effort, and Gemini 3 Pro with a high thinking level.
• Benchmarks without publicly available scores were re-evaluated under the same conditions used for Kimi K2.5.
• GPT-5.2 xhigh was not evaluated on all benchmarks due to service stability issues.

2. Text and Reasoning

• Various benchmarks evaluated with maximum completion budgets and averaging runs.

3. Tool-Augmented / Agentic Search

• Kimi K2.5 equipped with search, code-interpreter, and web-browsing tools for HLE with tools and all agentic search benchmarks.

4. Vision Benchmarks

• Details on max tokens, evaluation protocols, and score computations.

5. Coding Tasks

• Details on evaluation frameworks and scoring.

6. Long-Context Benchmarks

• Details on scoring and standardization.

7. Agent Swarm

• Details on step limits for main and sub-agents in swarm mode.

Original source