48 hours, one long-planned shockwave.
On July 18, 2026, Moonshot AI released Kimi K3. Within 48 hours, the global AI landscape showed visible cracks. On the night of July 19, Kimi published an announcement it never wanted to write: compute exhausted, new membership subscriptions suspended, all remaining capacity reserved for existing users. A model so hot it crushed its own servers — rare even in this industry.
In the markets, the first trading day after K3’s launch saw NVIDIA shed roughly $111 billion in market value in a single session; the Philadelphia Semiconductor Index fell 12.5% for the week, its worst in 15 months, and JPMorgan strategists called it “DeepSeek 2.0” outright. Anthropic’s reaction may be the most telling: it had planned to pull Fable 5 from subscriptions into pure usage-based billing — and the day after K3 launched, it reversed course, keeping Fable 5 in subscription plans permanently and issuing users compensation credits as goodwill. And David Sacks, the former US AI czar, posted that a Chinese model had taken first place in frontend coding for the first time, adding: at this rate, America loses the AI race. What did K3 actually do to trigger all of this?
Deconstructing K3: RL Scaling’s First Mass Deployment
Understanding K3 requires going back to 2024, when Yang Zhilin shared the K0-Math model internally and advanced a thesis: the next scaling law is Reinforcement Learning Scaling. The old scaling law was rote memorization — feed the model an ocean of data and it assembles the most-plausible answer at test time (next-token prediction). RL Scaling changes the assignment: don’t memorize answers; learn to drill, score yourself, find where you went wrong, fix it, repeat until correct — then write the process down as experience. K0-Math used exactly this logic to push past OpenAI o1 on mathematics. K3 is RL Scaling’s first large-scale extension from math reasoning to every scenario.
The technical report’s “Vision in the Loop” is that extension applied to visually-verifiable work: the model writes code, looks at the result, fixes what is wrong, looks again — a complete generate → inspect → iterate workflow. The demos backed it up: a 3D open-world game written from scratch (Three.js plus WebGPU), and a frame-precise launch trailer cut by the model itself from 56 source clips — transitions landing on the exact frame where the drumbeat hits hardest.
The hard numbers
K3 is the world’s first open-source 3-trillion-class model at 2.8T total parameters. Its KDA (Kimi Delta Attention) hybrid linear attention lifts million-token decoding 6.3×, and Attention Residuals add ~25% training efficiency at under 2% extra cost. Artificial Analysis scores K3 at 57 on its Intelligence Index — third globally behind Claude Fable 5 (60) and GPT-5.6 Sol (59), ahead of Claude Opus 4.8 (56), GPT-5.5, and GLM-5.2. On value: $0.94 per task, close to GPT-5.6 Sol’s $1.04 and half of Claude Opus 4.8’s $1.80.
K3’s Four Shockwaves
Wave one — users: peak load exposed real demand
Within 48 hours, requests far exceeded forecasts and pushed the cluster to its limits — evidence that demand for high-quality, low-cost models was massively underestimated. Kimi split its membership into Kimi core benefits and Kimi Code benefits to allocate compute more precisely. Meanwhile, foreign media reported Kimi has circulated a shareholder resolution seeking approval for a Hong Kong IPO within six months at a valuation possibly above $30 billion. ARR jumped from $100 million in March to nearly $300 million by June — tripling in a quarter, growing several-fold again after K3 — with API business now over 70% of revenue. The revenue structure has fully rotated to B2B.
Wave two — markets: the DeepSeek 2.0 chill
NVIDIA’s single-day $111 billion loss rhymes exactly with January 2025’s DeepSeek shock. Not coincidence: markets are repricing the destructive power of “Chinese open-source AI plus efficient compute.” When an open model delivers near-closed performance at a tenth the cost, the structural expectations for underlying compute demand crack.
Wave three — competitors: Anthropic’s overnight turn
Anthropic had planned to remove Fable 5 from subscriptions; the day after K3 launched, it committed to keeping Fable 5 in subscriptions permanently. A defensive response — when users discover K3 handles most Fable 5 tasks at a fraction of the price, the only play left is lowering the admission bar to keep the ecosystem.
Wave four — industry perception: the open gap shrinks from 6 months to 2-3
Berkeley professor and Databricks co-founder Ion Stoica put a number on it: “We used to say open models, especially Chinese open models, lag the frontier by six to nine months. Now the gap is maybe two to three months.” That judgment weighs more than any benchmark. Closed-model producers no longer own a time premium — and when anyone can complete nearly every AI task today, free, on open weights, the closed value proposition needs a new definition.
Where K3’s Boundaries Are
K3 is not flawless. The Intelligence Index is a composite; split it apart and real gaps show. On DeepSWE — long-horizon engineering in real codebases — GPT-5.6 Sol scores 73.0 and Fable 5 70.0 against K3’s 67.5: on long, multi-step, high-complexity engineering, K3 trails both. On Zerobench’s hard visual reasoning, Fable 5 scores 46.0 to K3’s 41.0 — Fable 5 can reconstruct a web app’s full source from a screenshot and extract precise data from dense scientific charts, still ahead on visual depth and precision.
Two deployment-level flaws also matter: harness dependency — K3 retains full thinking history, and if the caller fails to pass prior reasoning or the session switches from another model to K3, output quality drops noticeably; and over-eagerness — on ambiguous tasks it prefers deciding for the user over confirming intent, an asset in fast creative work and a liability in precision engineering flows. And one more: hallucination rates rose rather than fell — fatal for a model positioned for “long-horizon coding and knowledge work,” where one hallucinated decision point mid-12-hour task collapses the whole chain. Fable 5’s Constitutional AI keeps it more cautious — saying “I’m not sure” when unsure — which substantially cuts hallucination.
Back to the Start: Why Yang Chose This Road
A 36Kr detail rounds the story: after K3’s launch, Yang Zhilin’s CMU advisor Ruslan Salakhutdinov posted congratulations — and accidentally ignited a debate about why America couldn’t keep Yang Zhilin. Ruslan clarified: “If he wanted to stay in the US, he had many opportunities. But Yang was firmly determined to go back and start a company. He told me that if he didn’t at least try founding his own company, he would regret it for life.” Yang’s view: big-lab scientists, however capable, rarely get the resources and authority to ship products. “We saw our mentors hold industry titles, but essentially they were still doing research. Fundamental research matters, but it doesn’t break the wall between academia and industry.” From his PhD-year startup Recurrent AI in 2016 to Moonshot in 2023, the path never changed. K3 is RL Scaling validated from lab to product — and the company is his conviction that technology should become product, made real.
What K3’s 48 Hours Actually Revealed
Four structural shifts: First, RL Scaling is the new high ground. As pretraining’s marginal returns diminish, letting models learn to think via RL is the new differentiator — K3’s visual loop is one direction, not the end; whoever industrializes RL Scaling furthest owns the next iteration. Second, open models’ catch-up speed was underestimated. A 2-3 month gap means the closed moat is filling in fast; closed value increasingly rests on extreme long-horizon tasks, reliability, and safety guardrails — which most applications do not need. Third, compute infrastructure is the new bottleneck. K3’s subscription pause says good models do not lack users — they lack GPUs, straining model design, scheduling, and pricing alike. When capability overflows, the bottleneck moves from algorithm to physics. Fourth, global competition’s logic is shifting — from “who built the biggest model” to “who builds the most efficient, best-value, most accessible model.” An open Chinese model moving NVIDIA’s market cap and Anthropic’s pricing strategy is no accident; it is the new logic’s first public rehearsal.
For Practitioners
-
If your work involves long-horizon coding, K3 deserves a serious evaluation — the value proposition is real.
-
Watch harness consistency in deployment; switching to K3 mid-session from other models degrades output.
-
K3 excels in creative generation (games, content); pair it with verification mechanisms where reliability is critical.
-
Track RL Scaling. K3 is the route’s first scaled deployment — generations two and three will diverge far more.
References: 36Kr, “48 hours after K3’s launch: servers full, NVIDIA down, Anthropic changes subscriptions overnight” · Kimi K3 Technical Report · Artificial Analysis benchmarks
