What's New
A record of new features and key fixes across Kimi Code products — focused on the changes most worth knowing about.
kimi web runs in the foreground, kimi server deprecated
kimi web now runs in the current terminal foreground and opens the browser automatically; stop it with Ctrl+C. The old kimi server command tree is deprecated, its subcommands removed, with only kimi server kill retained as a fallback for stopping background services launched before v0.28.0. → kimi web
Thinking effort persists more conservatively
Thinking effort is now persisted only for levels below a model's top tier (max), so switching models no longer leaves you with an empty or unresponsive effort picker when the old max setting isn't supported. → Model configuration
Also in this release: the web model switcher now warns that switching models or thinking effort invalidates the existing prompt cache; the YOLO and Auto permission mode descriptions are corrected — YOLO auto-approves tool actions but may still ask questions, while Auto is fully autonomous and never asks.
/copy for the last reply
New /copy slash command copies the last assistant message to the clipboard — no more manually selecting long answers. → Slash Commands
Auto-refresh of the model list for API key users
When calling Kimi coding models with an API key, Kimi Code now automatically fetches the latest model list, so new models are usable without manual config changes. → Model configuration
Also in this release: OAuth connection failures now show the underlying network cause (DNS, refused connection, TLS, timeout) instead of a bare fetch failed; the built-in URL fetch tool's network safeguards are hardened so crafted domains and redirect chains can no longer reach loopback or internal network services.
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
Switching models invalidates the existing context cache, so consumption is higher right after switching; start a new session before using Kimi K3 for better results and lower consumption.
Coder sub-agent capabilities aligned with the main agent
The coder sub-agent now supports background tasks, todo lists, plan mode, skill invocation, and nested agents, so it can handle more complex coding workflows on its own. → Agents and Sub-Agents
Cache-invalidation hints in /model and /effort
Switching the model or effort now warns that the existing prompt cache will be invalidated, and suggests using /new to avoid extra token costs. → Slash Commands
Also in this release: the context-size indicator no longer under-reports the model's actual context usage; a resumed session without new activity is no longer incorrectly marked as just updated.
Attach any file in web chat
The built-in web UI now supports attaching files of any type in chat: drop a file anywhere on the window to send it, and files, images, and videos are shown as attachment tags in the message bubble. → Using Kimi Code CLI in IDEs
Aligned Anthropic effort configuration
Anthropic-compatible providers now apply the official effort configuration, with unknown models falling back to a 128k output limit and custom-named models getting correct thinking-strength controls for new sessions. → Providers and models
Also in this release: model request failures now show full diagnostics; two security issues were fixed — the web server's bearer-token check could be bypassed via percent-encoded API paths, and the session filesystem API could follow symlinks outside the workspace and access host files.
Built-in docs Q&A skill
A new built-in /check-kimi-code-docs answers Kimi Code product questions (CLI usage, configuration, membership, error codes) straight from the official docs, citing source links in its answers — so you no longer have to leave the terminal to look something up. → Built-in Skill Commands
Also in this release: the per-step LLM retry cap has been raised from 3 to 10, so transient provider-side failures (429 / overload) are retried automatically before the whole turn fails; and print mode (kimi -p) no longer exits after a single turn — as long as background tasks are pending it stays alive and feeds each result back to the main agent (print_background_mode now defaults to steer, and background tasks and subagents no longer time out by default).
Session export
The web UI can now export a session: run /export, or pick "Export session" from a session's overflow menu, to bundle the conversation and troubleshooting logs into a downloadable ZIP (up to 64 MiB) — handy for archiving or filing a report. → Exporting a session
Foreground commands background on timeout
When a foreground Bash command hits its timeout it is no longer killed — it moves to the background and keeps running, reporting its result when it finishes, so long-running commands are no longer lost midway. Set bash_auto_background_on_timeout = false under [background] to restore the kill-on-timeout behavior. → Built-in Tools · Shell
Also in this release: forking a session now fully preserves media attachments, plan files, background task output, and cron jobs, and a failed fork no longer leaves a partial copy behind; kimi web now defaults to the rebuilt agent engine.
⚡ HighSpeed is now available
Kimi Code now offers two tiers — Standard and HighSpeed — built on the same model with identical coding ability; HighSpeed delivers roughly 5–6× the output speed of Standard. Switch instantly with /model in the CLI, or set the model ID to kimi-for-coding-highspeed in third-party tools. HighSpeed requires an Allegretto plan or above. → How to Switch Models
Extra Usage is now available
When your subscription quota runs out, keep making requests with your Extra Usage balance — it bypasses the monthly / weekly / hourly membership quota limits, works whenever any limit is reached, only deducts the Extra Usage balance, and doesn't affect your membership quota refresh. Only subscribed members can enable it, at Kimi homepage → Settings → Membership Plan → My Quota. → About Extra Usage
Archived sessions
The web UI adds an Archived sessions page where you can browse and restore archived sessions in one click — open Settings → Archived to view them.
Reasoning preserved across turns
With Thinking enabled, Kimi models now keep their reasoning across turns by default so the model can build on its earlier thought; set [thinking] keep = "off" to disable it. → Model configuration
Compaction summary at a glance
After the context is compacted, the TUI shows a summary of what was compacted; press Ctrl-O to show or hide it within the compaction block — the same shortcut that collapses tool output. → Interaction and input
Also in this release: an experimental on-demand tool loader (select_tools) — with the tool-select flag enabled, capable models load MCP tools on demand instead of sending them all on every request, preserving the provider's prompt cache; and sessions that existed on disk but were missing from the list or returned 404 on direct access are fixed, as the server now rebuilds the session index on startup.
Automatic oversized-image compression
Images that exceed model limits are now automatically downsampled and re-encoded before reaching the model, cutting vision-token cost and avoiding provider image-size errors.
Model configuration overrides
A new [models."<alias>".overrides] table lets you override model metadata after provider-catalog refreshes — useful for pinning thinking-effort levels, default effort, and other fields so custom configs survive refreshes. → Model configuration
New web UI design system
The built-in web UI gets a refreshed design system with updated colors, typography, spacing, light and dark palettes, and subtler enter/exit and expand/collapse animations for a more consistent look and feel.
Also in this release: a new Cmd/Ctrl+K session-search palette filters by title, workspace, and recent prompt; consecutive tool calls are now grouped into a collapsible stack with diff line counts and inline previews for image, video, and audio results.
Plugins can provide slash commands
Plugins can now declare slash commands via a commands field in their manifest, registered as <plugin>:<command> and invoked with $ARGUMENTS expansion — turning common workflows into commands with no extra setup. → Plugin slash commands
Mermaid diagrams in web chat
Fenced mermaid blocks in assistant responses now render as diagrams in the built-in web UI, and KaTeX math and Mermaid parsing run in Web Workers to keep the UI responsive during live streaming.
Video input and Anthropic-compatible protocol
Kimi Code now supports the Anthropic-compatible protocol, and the ReadMediaFile tool can send video to the model as multimodal content — availability follows the current model's image_in / video_in capabilities. → Built-in tools
Web completion sound and notifications
The built-in web UI gains a completion sound, a completion notification, and an issue notification, each with its own toggle in settings. Issue notifications are off by default; only after you opt in is the issue text pushed to the desktop.
Custom outbound request headers
A new KIMI_CODE_CUSTOM_HEADERS environment variable lets you customize headers on outbound LLM requests (one Name: Value per line), and a User-Agent header is now sent to non-Kimi providers.
Also in this release: a provider 413 context overflow now recovers by compacting and retrying first, and compaction output is capped at 128k tokens by default to avoid provider max_tokens errors.
Shell mode
Type ! in the input box to enter shell mode and run terminal commands without leaving the conversation — output is written into the context for later turns. For long-running commands, press Ctrl+B to move them into the background. For example, run !gh auth login to sign in to the GitHub CLI without opening a new terminal. → Shell mode
Expose the web server to the internet
kimi web gains a --host option to expose the web server beyond your local machine, hardened with token authentication and rate limiting so you can reach the session UI from a remote box. → kimi web
Redesigned plugin manager
/plugins is now a single tabbed panel: Installed to manage installed plugins, Official for the Kimi-maintained marketplace, Third-party for marketplace plugins from other publishers, and Custom to install from a GitHub URL, zip, or local path — switch tabs with Tab / Shift-Tab. A confirmation prompt now appears before installing third-party plugins. → Plugins
Also in this release: fixed a startup crash on Linux caused by a native clipboard error, and /reload now refreshes plugin Skills so plugin changes take effect in the current session without starting a new one.
Extra workspace directories
You can now add directories beyond your working directory to a session: use /add-dir <path> to add one to the current session, or kimi --add-dir <path> at startup. When you choose to remember a directory, the path is written to the project-level .kimi-code/local.toml so it applies automatically in every session of that project. → Project-local configuration
Background tasks
Press Ctrl+B to move long-running foreground commands and sub-agents into the background, freeing up the session to continue, and view their status from the /tasks panel. → Session Management
Also in this release: provider safety-policy blocks are now surfaced instead of being silently treated as a completed turn, and the file-mention experience is improved.
Web session filtering and lazy loading
The web sidebar gains a session filter that matches on title and the most recent user prompt, and the chat view lazy-loads earlier messages as you scroll up — making long sessions easier to browse. → kimi web
Cap AgentSwarm ramp-up concurrency
A new KIMI_CODE_AGENT_SWARM_MAX_CONCURRENCY environment variable limits how many sub-agents AgentSwarm runs at once during its initial ramp-up, making large swarms less likely to hit provider rate limits. → Runtime switches
Also in this release: the web app no longer loads only the 20 most recent sessions, and web settings now display the current version.
Kimi Code Web Mode
A new web mode you can launch with kimi web or the in-CLI /web, continuing the current session in a browser chat interface. kimi web is an alias for kimi server run --open — it starts the local server in the background and opens the web UI automatically. → kimi web
Also in this release: when an OAuth token refresh still fails after internal retries, the underlying connection error is shown instead of a misleading re-login prompt; and kimi web failing to start in the background is fixed (v0.17.1).
kimi vis session visualizer
A new built-in kimi vis command launches the session visualizer in your browser, pointed at your local sessions, so you can see a whole session unfold at a glance. It supports --port/--host, --no-open, and kimi vis <sessionId> to deep-link straight to a specific session. → kimi vis
Also in this release: Anthropic-compatible providers no longer read ambient Anthropic shell credentials and custom headers, avoiding accidental credential leakage; and repeated compaction is fixed for cases where context stays over the blocking threshold.
All-sessions picker
A new all-sessions picker view lets you browse sessions across every working directory, search them by name, page through the list, and copy a ready-to-run resume command for sessions started elsewhere.
Reasoning follows your language
The same-language rule now extends to the model's reasoning, so its thinking follows the language you use while keeping code and technical terms in their original form.
Also in this release: MCP now supports legacy SSE servers alongside stdio and streamable HTTP transports (→ connection methods); and TUI components now wrap, compact, or truncate over-wide content to stay readable in narrow terminals.
Kimi K2.7 Code, our latest coding model, is now released and open-sourced — and available in Kimi For Coding. The new model takes effect only with Thinking On (thinking mode enabled).
Improved coding & agent performance over K2.6
+10.4% on Program-Bench, +11.4% on MCP Mark Verified, and +76.2% on SWE Marathon.
Reasoning efficiency
Less overthinking, with 30% lower reasoning-token usage compared to K2.6.
Long-horizon coding
Improved instruction following and higher end-to-end coding task success rates.
⚡ Kimi K2.7 Code HighSpeed Edition (beta coming soon)
The same model at roughly 5-6× the output speed.
Interrupt hook event
When you interrupt a turn (for example by pressing Esc), hooks now receive a dedicated Interrupt event — it fires in place of Stop, so external tooling no longer mistakes an interrupted turn for one that is still running. Programmatic interruptions such as timeouts don't trigger it. → Event reference
Kimi Datasource adds legal data
The official data plugin is now at v3.2.0, adding Chinese laws, regulations, and judicial cases — coverage now spans five data domains: stock market, macroeconomics, corporate registry, academic literature, and legal. → Plugin overview
Also in this release: images in tool outputs are no longer dropped when using OpenAI-compatible Chat Completions; since v0.13.1, running /undo without a count opens an interactive picker for choosing which prompts to roll back (→ Session management).
Custom color themes
Beyond the built-in dark and light palettes, you can now define your own as a JSON file — override only the colors you care about, the rest fall back to a base palette, and the file name becomes the theme name. Or run /custom-theme and let Kimi pick colors with you and write the theme file into ~/.kimi-code/themes/. → Custom themes
/import-from-cc-codex one-step import
Import selected instructions, Skills, and MCP settings from Claude Code and Codex — no need to reconfigure everything by hand when moving to Kimi Code. → Built-in skill commands
Also in this release: the marketplace list now flags installed plugins that have updates available — select one and press Enter to upgrade (→ Plugins); device login keeps the URL and code visible when the browser fails to open.
/swarm multi-agent parallel tasks
Use /swarm <task> to start swarm mode — multiple agents work on the same objective in parallel, with live progress display and rate-limit-aware automatic retries. Swarm mode exits automatically when the task completes. → Slash commands
Proxy support
The standard HTTP_PROXY, HTTPS_PROXY, ALL_PROXY, and NO_PROXY environment variables are now fully honored, including SOCKS proxies (socks5://, socks4://, and similar schemes). Loopback addresses always bypass the proxy, so local MCP servers on localhost continue to work after setting one. → Environment variables
Goals, background questions, and Sub-Skill discovery are now stable
These three features no longer require experimental opt-ins and are available by default. Micro compaction is also enabled by default — it trims older oversized tool results to reduce context overhead. To opt out, set experimental.micro_compaction = false in config.toml. (Removed in v0.22.1.)
Also in this release: Homebrew installations are detected and updated with brew upgrade kimi-code; Skills and global agent instructions load from KIMI_CODE_HOME when set; long shell commands in approval prompts now wrap to show the full command; multiple goal mode and ACP routing fixes.
Built-in Skills as direct slash commands
Built-in Skills (such as update-config) now appear directly in the slash command panel — no skill: prefix needed. Type / and they show up immediately, grouped ahead of external Skills. → Slash commands
Sub-Skill discovery (experimental)
A new experimental Sub-Skill system ships built-in sub-skill.review (audit existing Skills) and sub-skill.consolidate (merge Skills into hierarchical groups). Enable with KIMI_CODE_EXPERIMENTAL_SUB_SKILL=1. → Agent Skills
Also in this release: YOLO mode now asks for confirmation before starting a goal and suggests switching to Auto for unattended work; sub-agents show resume instructions on timeout; multiple goal queue bug fixes.
Goal queue
Use /goal next <objective> to line up upcoming tasks — when the current goal completes, Kimi picks up the next one automatically, no waiting, no manual trigger. Open /goal next manage to reorder the queue interactively. → Using Goals
Experimental features panel (/experiments)
Type /experiments in the TUI to open a visual toggle panel — flip a switch, confirm, and Kimi writes the change to config.toml and reloads the session. No manual config editing needed.
New built-in update-config Skill
You can now ask Kimi to edit its own configuration files directly.
This release also includes: /reload to hot-reload the session after config changes (→ Slash commands); kimi doctor to validate config file syntax (→ kimi doctor); Windows now fails early when Git Bash is missing.
kimi acp subcommand
Connect the CLI to your IDE via the ACP (Agent Client Protocol) — Zed, JetBrains AI Chat, and other editors can directly drive Kimi's sessions and tool calls without switching to a browser or terminal. → Use in IDEs
/btw side-channel conversation
Open a side conversation without interrupting the current main turn — ask a quick question or add context without affecting the Agent's main task. Entering /btw with no content opens the panel and waits for your input. → Slash Commands
Also: fixed Ctrl-G external editor invocation on Windows; unified TUI dialog and selector interaction.
Autonomous Goal mode (experimental)
Start with /goal <objective> and Kimi will work toward that goal across multiple turns until it's done or hits a decision point that needs you. → Autonomous goal
Background structured questions
When the Agent needs a decision from you, it parks that question in the background and keeps working on other steps — no more blocking on a single small choice. → AskUserQuestion
kimi provider subcommand
Add or remove providers directly from the terminal without entering the TUI — useful for scripts and CI. → kimi provider
Also: background auto-update is now on by default (disable in tui.toml); context compaction (/compact) now attaches the to-do list to the summary so the Agent doesn't forget where it left off.
/provider interactive provider manager
Type /provider in the TUI to open a visual interface for viewing, adding, and removing providers — no more manually editing config files to switch models. → Platforms & Models
KIMI_MODEL_ADAPTIVE_THINKING environment variable
When connecting to a custom endpoint, force-specify whether adaptive thinking is enabled instead of relying on model-name inference. → Model fields
Also: scheduled task trigger times now display in local timezone.
KIMI_MODEL_* environment variable channel
Set a few environment variables to temporarily switch models without touching config files — changes expire on restart, ideal for testing or CI. → Define a model via environment variables
Install plugins directly from a GitHub URL
Paste a repo link to install community plugins, with support for pinning specific versions. The plugin manager labels each install with a trust level. → Installation & Management
Removed default step limit
The per-turn 1000-step cap is gone — long tasks no longer get force-interrupted. → loop_control
Scheduled tasks
Set timed schedules or one-off reminders in natural language — Kimi executes automatically at the right time without you having to watch. → Scheduled Tasks
/auto permission mode
A more restrained automation mode than /yolo — tool approvals are handled automatically, but the Agent won't ask you questions. Good for unattended runs where you don't want to fully open the gates. → Interaction & Input
Plugin system launched
Install plugin packages that bundle Skills and MCP servers. The official marketplace is live, with Kimi Datasource as the first published plugin. → Plugins
Session export
Export conversation history as a Markdown file, or package it as a ZIP to submit feedback. → Sessions & Context
Permission system redesign
Read-only operations outside the working directory no longer trigger approval prompts — fewer unnecessary interruptions. → Permission config
OpenAI-compatible reasoning models work out of the box
DeepSeek, Qwen, One API, and other models with thinking tokens need no extra configuration. → openai provider
/logout with provider selection
A provider picker now appears on logout to avoid accidentally logging out of the wrong provider.
/connect command
Search a public directory of providers and configure one in a single step — no need to write config files by hand. (Superseded by /provider in v0.7.0.)
Kimi Code CLI — initial release. The entire codebase has been rewritten from Python / uv to TypeScript / Node.js. This isn't an incremental update — it's a new foundation, with a redesigned architecture, installation flow, and configuration format.
Key differences from the legacy kimi-cli:
| Legacy kimi-cli | Kimi Code CLI | |
|---|---|---|
| Runtime | Python + uv | Node.js (no Python dependency) |
| Install | uv tool install | One-line curl script or npm install -g |
| Config file | ~/.kimi/config.toml | ~/.kimi-code/config.toml (incompatible format) |
| Terminal UI | Basic text output | Full TUI (chat view + status bar + approval panel) |
| Startup speed | Slower (Python cold start) | Faster (Node.js native binary) |
| Multi-provider | Limited | Built-in Anthropic / OpenAI / Gemini / Vertex, and more |
| Sub-agents | Not supported | Built-in coder / explore / plan sub-agents |
| Plugin system | Not supported | Supported (since v0.4.0) |
| Scheduled tasks | Not supported | Supported (since v0.5.0) |
Existing kimi-cli data (config, sessions, MCP declarations) can be migrated in one step with kimi migrate. → Version Upgrade