AI news

Moonshot launches Kimi Hosted Agents, a managed cloud harness for long-running agents

Moonshot says Kimi Hosted Agents is now generally available. It turns the agent harness—code execution, context and sessions, tool calls, permissions, and interrupt/resume—into a managed cloud service, so teams can run long-horizon and multi-agent jobs without building that stack. Each agent gets an isolated sandbox with network access controls; operators can set boundaries, stop a session, and export audit trails. Usage billing and reports are included. The platform ships more than 140 skills and 30-plus plugins, and teams can add their own tools and data. An optional Inference Hook can check inputs and tool outputs against enterprise rules before inference; it is limited to verified enterprise accounts for now. Docs include eight finance agent templates. Sign-up is at platform.kimi.com.

Why it matters Getting agents into production often means months of harness work before any business logic; a managed runtime shifts that burden off the team, without changing the model.

Kimi

Moonshot AI opens Kimi K3.1 Pro weights: 2.8 trillion parameters, 104 billion used per token

Official source · auto-listedOriginal title moonshotai/Kimi-K3.1-Pro

Moonshot AI has made the full weights of Kimi K3.1 Pro public on Hugging Face, so anyone can download and deploy them, delivering on its earlier pledge. The model and its API went live on October 24; what is new is the open weights. The model has about 2.8 trillion parameters in total but uses only about 104 billion for each token. It supports a 1-million-token context window and accepts both images and text. It is Kimi K3 further trained with large-scale reinforcement learning, aimed at coding, 3D modeling and financial research. The license is the "Kimi K3.1 Pro License". Self-reported by Moonshot: K3.1 Pro scores 65.7% on Terminal-Bench 4.0. Moonshot also says it removed cyberattack-related data from training for the open release to reduce misuse risk.

Why it matters A trillion-scale model whose weights anyone can download and run on their own hardware gives developers another strong self-hostable option for coding and agent work.

Kimi

Moonshot apologizes for Kimi K3.1 Pro rate limits, resets 25% of every subscriber's monthly quota

Moonshot AI's Kimi apologized on Oct 26 for rate limiting and unnecessary quota use on its K3.1 Pro subscription plan, calling the issues its own. Every active subscriber gets a one-time reset restoring 25% of the monthly quota (about a week of use); usage below 25% resets to zero. Kimi also says it is cutting unnecessary sub-agent calls and output length.

Why it matters K3.1 Pro launched on Oct 24, and two days later Kimi had to publicly compensate subscribers, a sign of the compute and cost strain of serving a very large model on flat plans; fewer sub-agent calls may change long-task behavior.

Kimi

Kimi Code 2.3.2 makes single-agent the default; subagents and Agent Swarm now need explicit opt-in

Kimi Code CLI 2.3.2 changes its defaults: the agent now handles tasks itself instead of proactively spawning subagents, and delegation must be requested in the prompt, AGENTS.md, a skill or a multi-agent mode. The agent can no longer launch AgentSwarm on its own; swarms run only when the user turns on swarm mode, so they are unavailable in headless kimi -p runs and ACP clients such as Zed.

Why it matters Automatic fan-out can turn one prompt into many concurrent model calls. Opt-in delegation makes token use more predictable, but workflows that relied on automatic swarms need explicit instructions.

Kimi

Moonshot launches Kimi K3.1 Pro and deliberately weakens its hacking ability

Moonshot AI released Kimi K3.1 Pro, a post-trained K3 with a 1M-token context, on Kimi.ai, Kimi Work, Kimi Code and the API, with open weights promised by Nov. 4. It reports 65.7% on Terminal-Bench 4.0 and an internal hallucination rate cut from 23.4% to 8.3%. It removed offensive-security training data: on xdev-bench-hard the release scores 81.0% versus 90.5% unfiltered.

Why it matters Moonshot is openly trading away some capability to reduce offensive cyber skills, addressing a core worry about releasing open weights. All benchmark numbers are self-reported and not yet independently verified.

Kimi

Tencent's WorkBuddy adds Moonshot's Kimi K3.1 Pro, including a faster Highspeed variant

Tencent's workplace agent WorkBuddy said on Oct 24 it now offers Moonshot AI's new Kimi-K3.1-Pro with a 1M-token context window, plus a Highspeed variant Tencent says matches its intelligence at roughly 2-3x the output speed. Citing tight capacity, access goes first to enterprise flagship and individual paid subscribers. The speed figures are Tencent's own claim.

Why it matters Kimi K3.1 Pro landing in a Tencent product days after launch shows large platforms such as Tencent will ship rivals' models in their own apps, while the tiered rollout points to the compute cost of long-context inference.

Kimi