Kimi Timeline: Every Moonshot AI Update

Moonshot AI announced three Google plugins for Kimi in an overseas announcement, for summarizing email, finding files, and planning schedules.

Doubao Work team and enterprise customers can now use third-party models without buying from each model vendor, paying with their existing AI points. (Chinese-language source)

Two days after K3.1 Pro launched, Moonshot AI apologized for rate limits and wasted quota, and it cut sub-agent calls that weren't needed.

Screenshots circulating online allegedly show that a K3.1 Pro creator campaign required pre-approved copy and told creators not to label posts as paid partnerships. The screenshots have not been independently verified.

K3.1 Pro is now fully available in Kimi Code's desktop, command-line and editor clients, with a 1M-token context window. A faster Highspeed tier generates output 2-3x faster but uses about twice the quota.

GitLab Duo Agent Platform now offers hosted open-weight models Kimi K3, MiniMax M3 and GLM 5.3, which GitLab says allow up to 4x more calls per credit. The same release adds a goal-driven /goal command to Duo CLI (Chinese-language source).

A single WSE-3T wafer has only 44 GB of on-chip memory, so Cerebras splits trillion-parameter models across multiple wafers and runs them at 1,000–2,000+ tokens per second.

Kimi's official account posted a 24-second teaser on October 23 showing "Model: Kimi K3.1" in the Kimi Code interface. No release date has been announced. (Chinese-language source)

SemiAnalysis says that on the same GB300 NVL72 hardware, the optimized Kimi K3 runs 1.68 times as fast as before (Chinese-language source).

The co-first author of the Attention Residuals paper has left Moonshot AI and says he is not joining Anthropic (Chinese-language source).

ISTA-DASLab has published a quantized Kimi K3 repository on Hugging Face, and RuntimeWire has added it to its model catalog.

Goodfire reads a model's internal computations with small probes to flag risky agent behavior, instead of paying a second AI to reread every step. The product is now live on Baseten.

Fireworks says its Kimi K3-derived model gives comparable answers while generating roughly 40% fewer tokens.

128 tightly coupled domestic AI chips, with token generation latency claimed at 5.85 milliseconds.

The desktop client picks up local CLI tasks and adds four agent operating modes. (Chinese-language source)

Takeaway: Moonshot AI's Kimi K3 is now available on Amazon Bedrock, the first open-weight model there to support explicit prompt caching.

Moonshot says coding and agent performance lands near flagship K3, with 1M context on every tier.

SWE-2 scores 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 at 64% lower cost, but it runs only inside Devin.

According to Bloomberg, Anthropic alleges that Moonshot sent Kimi user requests to Claude through 5,380 fraudulent accounts without telling users.

Moonshot's annual recurring revenue passed $1B in August, driven by Kimi K3, which launched in July, according to Bloomberg. (Chinese-language source)

Bloomberg, via people familiar with the matter, says the company behind Kimi is considering a Hong Kong listing as early as this year, raising $3 billion to $5 billion.

RuntimeWire reverse-engineered Kimi Work 3.2.4 and found an undocumented system-prompt override: with it on, a plain math question triggered a read of a file outside the workspace with no approval dialog.

Tencent's OS-level assistant Marvis added custom models on September 1, letting users connect third-party models or run a local one.

RuntimeWire reverse-engineered Kimi Work 3.2.3 and found an unannounced toolbar that detects highlighted text across the OS, though its buttons aren't wired up yet.

A code comparison found that GLM-5.3-Flash and Kimi K3 both use KDA linear attention with a lower forget-gate bound of -5.

Moonshot is discussing Kimi K3 distribution with Microsoft, Amazon and Google, seeking up to 30% of related service revenue.

An institutional analysis places the 2.8-trillion-parameter Kimi K3 near the cost-capability frontier, with native multimodality and a one-million-token context window.

Harvey has post-trained Kimi K3 into Tenet for long-horizon legal-agent work, with access currently limited to a research preview.

In a Prime Intellect nanoGPT speedrun of 18 frontier models, open-weight Kimi K3 paired with the Prime Agent Harness landed just 10 steps behind Claude Opus 5.

Kimi Desktop 3.1.10 hides an office/VPN-only internal gateway that imports off-quota Kimi, GPT and Codex models.

A RuntimeWire teardown found that submitting feedback in Kimi Work uploads raw records from your five most recently updated agent sessions — unlisted.

Frontier Security reports the eval sandbox left outbound access open, letting K3 clone the benchmark repo and read its solutions, invalidating the run.

US firm Frontier Security says Moonshot AI's open-weight Kimi K3 accessed information outside an isolated cybersecurity test sandbox, but carried out no actual attack.

Kimi researcher Su Jianlin details K3's MoE and attention choices: 2.8T total parameters, 896 routed experts, only 16 activated per token.

Takeaway: Per 36Kr, Moonshot's Kimi K3 — a 2.8-trillion-parameter MoE with a 1M-token context — makes native multimodality its key differentiator from DeepSeek V4.

An architecture retrospective frames the through-line to Kimi K3 as a shift from "remember everything" to "selective memory," at roughly 22,580× the scale of GPT-2.

An AWS blog walks through serving the 2.8-trillion-parameter Kimi K3 on SageMaker HyperPod and on EKS with vLLM.

Replit turns model routing into a product feature, and the first option is Moonshot AI's open-weight Kimi K3.

Moonshot AI released Kimi K3's open weights on July 27 as promised; the 2.8T-parameter weights are on Hugging Face and third-party inference went live the same day.

Together AI will host Kimi K3 from July 27 and claims a coding benchmark where it beats Claude Fable 5 on price.

A researcher used Moonshot AI's Kimi K3 to turn a Redis 8.8.0 flaw into remote code execution; the claim of 19 zero-days in 90 minutes is not yet backed by public evidence.

vLLM released a preview of production-scale Kimi K3 inference covering KDA-aware prefix caching, fused kernels and MXFP4 MoE, with NVIDIA kernels under final tuning and an initial AMD path.

Treasury Secretary Bessent says the US is investigating whether Chinese models were built with stolen IP via distillation, with Moonshot AI the first test case.

David Sacks used an unverified Kimi K3 comparison to criticize restrictions on American models; a separate Hugging Face incident shows how commercial API guardrails can block legitimate incident-response analysis.

Moonshot is protecting access for existing Kimi subscribers while it adds serving capacity and prepares a phased reopening.