Search

Moonshot AI apologizes for K3.1 Pro rate limits and restores 25% of the monthly quota, Meta releases InteractiveWorld, Kimi Code 2.3.2 no longer launches sub-agents on its own

JLS42 /
Article generated by artificial intelligence
kimi moonshot agents coding-tools google gemini gemini-cli meta recherche huggingface elevenlabs audio ia-generative actualites
Moonshot AI apologizes for K3.1 Pro rate limits and restores 25% of the monthly quota, Meta releases InteractiveWorld, Kimi Code 2.3.2 no longer launches sub-agents on its own

ai-powered-markdown-translator

Article translated from French to English with gpt-6.1-sol.

View project on GitHub ↗

Two days after launching Kimi K3.1 Pro, Moonshot AI apologizes for rate limits and unnecessarily consumed quota on its subscription plan, and restores 25% of each subscriber’s monthly quota. In brief, Kimi Code 2.3.2 no longer launches sub-agents on its own initiative, the Gemini CLI nightly arrives with no changes, Meta releases the InteractiveWorld simulator, and Hugging Face posts introduce the Mundo world model, the A5 personal agent and Kos, a Polish voice assistant. Lastly, ElevenLabs adds two figures to its report on the agent that calls Cadence patients.

Kimi K3.1 Pro: Moonshot AI apologizes and restores 25% of each subscriber’s monthly quota

October 25 — Two days after launching Kimi K3.1 Pro, Moonshot AI apologizes to its subscribers.

We’re sorry for the rate limits and unnecessary quota consumption on the K3.1 Pro subscription plan. These were our issues, and we should have fixed them sooner. — @Kimi_Moonshot on X

As compensation, every subscriber, regardless of plan, receives a one-time credit of 25% of their monthly quota, a week’s usage according to Moonshot AI; usage below 25% is reset to zero. The plan has been adjusted to launch fewer unnecessary sub-agents and produce shorter outputs, with no figures or new pricing announced.

Earlier that morning, Kimi Code 2.3.2 had stopped launching sub-agents on its own (see Briefs): a coincidence in timing, as the tweet does not mention the coding agent.

Briefs

Coding agents

  • Kimi Code 2.3.2 — The agent now handles tasks itself instead of launching sub-agents on its own initiative: delegation must be requested in the message, in AGENTS.md, in a skill or through a multi-agent mode. Agent Swarm now starts only when swarm mode (swarm) is enabled, so it no longer runs in kimi -p or ACP clients such as Zed. 🔗 source
  • Gemini CLI, October 25 nightly — It contains no changes: the same commit as the October 24 nightly, only the version numbers change, no PR has been merged since the evening of October 23, and stable v0.65.0 and preview v0.66.0-preview.0 remain in place. 🔗 source

World models and simulators

  • InteractiveWorld — Catch-up from October 24: Meta releases a simulator where an assistant sees through the eyes of a simulated person and speaks to them, across 44 locations drawn from BEHAVIOR-1K, alongside a benchmark of 120 multi-turn tasks checked against the world state. On its project page, Claude Fable 5.1 achieves a 41.7% task success rate, the eleven other models tested achieve 33.3% or less, while an oracle assistant succeeds on every task; non-commercial CC BY-NC 4.0 license. 🔗 source · 🔗 project page
  • Mundo — Catch-up from October 23: in a Hugging Face post, Yihang Luo, Chen Change Loy and their coauthors introduce a world model that reconstructs a room from a few photos, simulates its objects and hands rendering over to a video model: changing the viewpoint, opening a drawer, bringing in a robot. Still slow according to the authors: ten to twenty minutes for reconstruction, rendering at around 1.5 frames per second on eight A800 GPUs; code and weights are slated for late November. 🔗 source · 🔗 repository

Personal assistants

  • A5 — Catch-up from October 23: Honglin Guo, Tao Gui (SII) and their coauthors introduce a proactive personal assistant that reuses past preferences and decisions, monitors connected sources and suggests next steps on its own. Its first building block, A5 Harness, adds memory across sessions to a ReAct loop; the code has not yet been published, but a technical preview can already be installed on macOS and Linux. 🔗 source
  • Kos — Mateusz Piesiak releases a Polish voice assistant that is allowed to answer “I’m not sure”: in a Hugging Face Space on ZeroGPU, with no pay-per-use API, Whisper large-v3 listens, the open model PLLuM-12B responds, the code itself handles weather, exchange rates and arithmetic, and every free-form response is checked before being spoken. 🔗 source

Voice and healthcare

  • ElevenLabs and Cadence — Catch-up from October 21: in a webinar recap, ElevenLabs adds two figures about the ElevenAgents agent that calls patients monitored remotely by Cadence whenever a reading falls outside the thresholds: 60% of critical calls answered compared with 41% previously, and patients picking up 47% more often for the agent than for human callers. 🔗 source

What this means

A quota-based subscription promises a certain amount of work, and an agentic model can use it up without the subscriber realizing: rate limits, sub-agents launched unnecessarily, overly long outputs. Moonshot AI acknowledges that these problems were its fault, compensates for them with quota, returning 25% of the month on every plan, and adjusts the plan settings without announcing new pricing or quantifying those adjustments. Subscribers will see on their usage bar whether their quota really lasts longer.

At Moonshot AI, the agent does less on its own. On the K3.1 Pro plan, the provider reduces sub-agent calls it considers unnecessary; on the same day, Kimi Code 2.3.2 delegates only on request, and Agent Swarm, introduced in February and expanded with K2.6 in April, now starts only in swarm mode. Moonshot AI does not link the two decisions, but they move in the same direction: fewer sub-agents launched by default.

Control also extends to the response itself. Kos is allowed to answer “I’m not sure” and checks every free-form response before speaking it; at Cadence, an alert is closed only if every symptom question receives a clear “no,” with “not sure” counting as a “yes,” and a second judge checks the conclusion against the transcript; InteractiveWorld judges an assistant by the state of the world, not by what it says.

Sources