中/EN
Frontline Hotspot
Frontline Hotspot

Kimi K3 Goes Open Source Tonight: 2.8 Trillion Parameters, World's Largest, As Yang Zhilin Closes the China-US Model Gap to 3 Months

On the evening of July 27, Moonshot AI open-sourced Kimi K3's weights: a 2.8-trillion-parameter MoE with a 1-million-token context, the world's largest open-source model, benchmarking against Anthropic's Fable 5. From the July 16 API launch to tonight's weight release, Yang Zhilin used ten days to compress the China-US model gap from 6-9 months to 3-5. Breakdown of specs, benchmarks, the comeback story, and the White House accusation.

Published July 27, 20263 min read
<!-- kimi-k3-open-weights-opensource | hotspot | Kimi K3 Open Source Tonight -->

On the evening of July 27, Moonshot AI posted Kimi K3's model weights to its download page, free, commercially usable, and self-hostable. It is the world's first open-source model to cross the two-trillion-parameter mark while letting developers freely download and modify it. From the API and app launch on July 16 to tonight's weight release, only ten days have passed.

How Big Is 2.8 Trillion Parameters

K3 has 2.8 trillion total parameters, built on Stable MoE with 16 of 896 experts active per layer. It has a 1-million-token context window, native multimodal vision, and a built-in always-on "thinking mode." For comparison, DeepSeek V4 Pro is about 1.6T, Xiaomi 1.02T, and Alibaba 397B. K3 is roughly 75 percent larger than the runner-up. Two in-house architectural moves stand out: Kimi Delta Attention (a hybrid linear attention mechanism) and Attention Residuals (a drop-in replacement for residual connections), both previously published on GitHub. Weights ship in MXFP4 with MXFP8 activations, and quantization-aware training was baked in.

Benchmarks Against Fable 5

Moonshot's official benchmarks call K3 "competitive" with Anthropic's Fable 5, and "substantially outperforming" Opus 4.8, GPT-5.6 Sol, and GPT-5.5. Independent trackers are more measured. Artificial Analysis ranks it third overall, behind Fable 5 and GPT-5.6 Sol Max. Arena's front-end coding blind test puts it first, ahead of the top US models. Vals AI ranks it second. The catch is cost: Fable 5-class reasoning at a fraction of the price.

Yang Zhilin's Comeback

Founder Yang Zhilin is 34, a Tsinghua and Carnegie Mellon alum, and a Pink Floyd fan. The company name (月之暗面, "dark side of the moon") comes from the album. For 18 months the company was pressed by DeepSeek and its market position eroded; K3 is the counterpunch. The numbers are stark. Post-K3 daily revenue grew sixfold, June ARR hit $300 million (up from $200 million in April), and the company is seeking a new round at a $50 billion valuation, with a Hong Kong IPO possibly this year. On launch day, competitors' stocks dove: Z.ai down 30 percent, MiniMax down 16 percent, Alibaba down 4 percent.

The Math Behind Open-Sourcing

Why not stay closed and charge? Yang's logic is to trade openness for a user base and developer community, and vie to be the gravity center of global open-source AI. The geopolitical edge is sharper. Reuters noted that open-sourcing lets Chinese firms both flex muscle and expand influence, neatly countering US chip restrictions. The White House is not idle. Office of Science and Technology Policy director Michael Kratsios publicly accused Moonshot last week of training K3 on banned Nvidia chips and running large-scale distillation against US models including Fable; Moonshot has not responded. Researcher Nathan Lambert's read: the open-versus-closed, China-versus-US gap has compressed from six to nine months down to three to five.

Stacking parameters to 2.8 trillion is engineering capability. Daring to dump all the weights into the open at the peak is a strategic choice. The download page is live; whether to run it and how is up to your compute.

References

This article is AI-assisted and human-edited. Last updated: 2026-07-27

FAQ

What is Kimi K3? When was it open-sourced?
Moonshot AI's flagship model released July 16, 2026-a 2.8-trillion-parameter MoE with a 1-million-token context. On the evening of July 27, the weights were opened for free download, commercially usable and self-hostable, currently the world's largest open-source model.
What does 2.8 trillion parameters mean? How do regular users access it?
Parameters measure an LLM's scale; 2.8 trillion makes K3 the largest open-source model, about 75% bigger than DeepSeek V4 Pro. With weights open, developers can download, modify, and self-host-needs ample compute. Regular users just use the Kimi app or API, no need to touch the weights.
Why is Kimi K3 said to benchmark against Anthropic's Fable 5?
Moonshot's official benchmarks call K3 "competitive" with Fable 5, substantially outperforming Opus 4.8 and GPT-5.6 Sol; independent tracker Artificial Analysis ranks it third (behind Fable 5 and GPT-5.6 Sol Max), and Arena's front-end coding blind test ranks it first. But its cost is far below Fable 5.

Related

Frontline Hotspot

Kimi K2.8 Preview: Chasing the Coding Agent's Daily Traffic

On 2026-09-14 Moonshot AI shipped Kimi K2.8 Preview, its new primary model, rolling out across Kimi Code and Kimi Work. Officially its overall performance approaches the flagship K3, coding and agent ability improved across the board over K2.7 Code, and thinking efficiency improved notably. It supports three reasoning-effort levels, low, high and max, aligned with K3 and defaulting to max, switchable in the CLI with /effort, plus image and video input. The 1M-token context window is open to every membership tier including the free Adagio tier. The model ID stays kimi-for-coding, so Claude Code, OpenCode, Codex and other third-party tools upgrade with zero configuration changes, and Allegretto tier or above can switch to kimi-for-coding-highspeed for speed. At the routing layer, requests to K3 with thinking disabled are handed to the no-thinking K2.8 variant. Billing rides the subscription, from 0 to 699 RMB per month, not per token. Moonshot's ARR grew from 300 million USD in June 2026 to over 1 billion USD in August. This piece skips the launch recap: its core claim is that the bundle is aimed at capturing the daily traffic of coding agents, shifting competition from benchmark scores to per-task cost and migration friction. It also names three things still unverified: no official benchmarks at all (the widely circulated 63.2 percent SWE-bench Pro and 81.2 percent OSWorld-Verified figures belong to Claude Sonnet 5, and are labeled as such here), weights not yet announced as open, and no dedicated GitHub repository for K2.8, since the moonshotai organization's 43 repos stop at K3, K2.5 and K2, which per our editorial rule makes it a hotspot topic only, not an open-source write-up.

Sep 15, 20269 min read
Frontline Hotspot

DeepSeek V4.1 Flash Open Weights: The Asymmetric Design

DeepSeek open-sourced V4.1 Flash on 2026-09-10: a 552B-parameter MoE with an asymmetric Causal-Encoder-Decoder design that activates only 8B on input and 16B on output, natively multimodal, with officials citing significant KV Cache compression to cut agent-scenario cost. The API shipped alongside it - just switch the model name to deepseek-flash - and Tencent WorkBuddy, CodeBuddy plus OpenCode have integrated it fully. The model first surfaced on 9-08 as an internal preview build before being promoted on 9-10, a timeline worth noting in itself. This piece breaks down each release claim, argues the real engineering signal is not parameter count but the shift in long-context and agent cost structure implied by the asymmetric design plus 8B input activation, runs the numbers on what KV Cache compression means for accumulated multi-turn trajectories, and closes with cold takes: no published benchmark comparison, an unresolved relationship to its own V4-Flash, and concurrency and pricing still unconfirmed. Note that what shipped is model weights on HuggingFace; there is no dedicated code repository for V4.1 Flash under the official DeepSeek org.

Sep 10, 20269 min read
Frontline Hotspot

August 31, 2026: Sonnet 5 Reprices, GPT-5.4 Exits Codex, and Two Kimi Models Sunset on the Same Day

Three unrelated events landed on the same date. Claude Sonnet 5's API launch pricing expired, moving from $2 input and $10 output per million tokens to the standard $3 and $15. GPT-5.4 and GPT-5.4 mini stopped being offered to Codex users signed in with a ChatGPT account, replaced by GPT-5.6 Terra and Luna. Moonshot AI sunset kimi-k2.5 and moonshot-v1 on the same day, with kimi-k3 as the stated migration target. The layer most people miss is the second one: Sonnet 5 also changed tokenizer, so the same input now maps to 1.0x to 1.35x more tokens, which compounds with the rate change to a 65% to 100% real increase on coding workloads and close to 50% on plain text, while leaving Consumer subscriptions untouched. This piece splits the three events into sunset, replacement, and repricing, gives the urgency and response for each, and closes with a two-month expiry calendar: Claude Code limits on 2026-09-14, DashScope retiring 30-plus model IDs on 2026-10-10, deepseek-chat and deepseek-reasoner deprecated on 2026-10-24, OpenAI leaving Cursor on 2026-11-12, and the GPT-5.6 Sol promotion ending around 2026-11-21. Prices are a 2026-08-31 snapshot and source disagreements are flagged inline.

Aug 31, 20266 min read