Kimi K3 is not just another large model — it is the model that raised prices. On the evening of July 16, 24 hours before WAIC opened, Moonshot AI quietly released K3: 2.8 trillion parameters, the world’s first trillion-scale open model, beating GPT-5.6 and Fable 5 on multiple benchmarks.
But what truly shook the industry was not the performance. It was the pricing: K3 charges 20 yuan per million uncached input tokens and 100 yuan per million output tokens — 20× DeepSeek V4 Pro. A Chinese AI company openly raising prices twenty-fold in a market where everyone is shouting “cut prices” is not just a product launch. It is an industry signal.
K3’s Three Layers of Breakthrough
Layer 1: architecture — KDA hybrid attention
K3’s technical core is its self-created KDA (Kimi Delta Attention) hybrid linear attention, paired with Attention Residuals. The principle is not complicated: run full attention at the critical nodes of a task to guarantee precision, switch to linear attention on non-critical paths to cut inference cost. The effect is dramatic — decoding on million-token contexts up to 6.3× faster, training efficiency up roughly 25%, at under 2% additional cost. This is the technical validation of the “stack parameters” route: 2.8T parameters is not mindless accumulation but architecture innovation making “big” usable. K3 answered a fundamental industry question with data — the ceiling on parameter scale is far from reached; the question is who can afford to run it.
Layer 2: paradigm — from conversational AI to project-based AI
K3 fundamentally serves three scenarios: long-horizon coding, end-to-end knowledge work, and deep reasoning. It is not a chatbot; it is a project executor. Engineer feedback bears this out: on complex programming tasks, K3’s one-shot code pass rate exceeds OpenAI’s comparable models, and it completes cross-file, engineering-grade changes in a single session — a million-token context holds an entire codebase. That is not incremental improvement; it is a leap in the human-machine interaction paradigm. Users stop “talking with AI” and start “handing tasks to AI.” Yang Zhilin has long named Kimi’s three thrusts — token efficiency, long context, agent swarms — and K3 integrates them into one continuum.
Layer 3: business — from price war to capability pricing
K3’s pricing is the most consequential part of the story. Moonshot’s head of enterprise business answered critics with a pointed line: open models and Chinese models should not wear a cheap label — we built an SOTA-level model, and it deserves rational commercial pricing. Behind it sit three judgments:
-
Capability has clear tiers — not every task deserves K3, but high-value tasks deserve the premium;
-
Total-cost thinking — K3 needs fewer rounds and less rework on complex coding, so the all-in cost may not be higher;
-
Global positioning — K3 prices at roughly 60% of GPT-5.6 Sol and 30% of Fable 5, which is competitive for overseas buyers.
Goldman Sachs’ research note put a label on it: this marks a brand-new node in Chinese models’ march toward pricing power.
Three Choices Facing Chinese AI
After K3’s launch, four US tech giants’ shares sagged; on Hong Kong’s board Zhipu fell 28.49% in a day and MiniMax 15.62%. An open model release triggering global capital-market chain reactions means we are crossing a structural inflection. Three choices now define the landscape:
Choice one: open or closed? OpenAI’s future-strategy head Dean Ball calls open source “decelerationism” that erodes commercial AI’s return on investment. Yang Zhilin’s doctoral advisor Russ Salakhutdinov counters that Kimi’s open release is “remarkable and critically important.” Russ’s observation cuts to the bone: open weights are always the harder road, because competitors can take the model, study it, use it to generate data, and train their own. But the ecosystem dividend is enormous — the community will distill and optimize K3, compressing the 2.8T giant into smaller, more efficient models. Open source is ecosystem efficiency fighting single-model efficiency.
Choice two: parameters or efficiency? K3 needs roughly 20 H100s at runtime and can spend over 30 minutes on complex coding tasks; within 48 hours of launch it suspended new registrations for lack of compute. That is not unique to K3 — it is the necessary price of the parameter route. But Russ’s predicted path is clear: Kimi will build a tiered model family like every other lab — one large, one medium, one small. Once the foundation model hits SOTA, efficient operation is comparatively easy.
Choice three: low prices or pricing power? The industry’s past year was a price war — DeepSeek, Zhipu, and Alibaba competed API costs toward zero. K3 tore open another road: if you hold a genuine technical lead, you can decline to be hostage to “cheap.” That is a lesson for the entire Chinese AI industry. When everyone is trapped in the doom loop of cut prices → lose money → invest less → fall behind, K3 offers the virtuous circle: good product → high pricing → more investment → better product.
Global Competition Enters the Capability-Tier Era
The structural change is bigger than one company. For Moonshot itself: on July 22, reports said the company plans a final pre-IPO round in August targeting a $50 billion pre-money valuation — K3’s technical showing makes the number defensible. For Chinese AI: K3 proves Chinese teams can build world-class models and walk their own technical line — not following OpenAI, but differentiating on architecture. For the global landscape: when a trillion-scale open model is free to download and deploy locally, the whole industry’s capability floor rises; startups access top-tier capability, and competition shifts from “who can use the best model” to “who can build the best product with it.” For pricing: multi-model strategy becomes the mainstream — enterprises stop choosing between “expensive and great” and “cheap and adequate,” instead routing each task to the optimal model. Routing layers like OpenRouter gain value accordingly — which is precisely the logic behind Stripe’s $10 billion acquisition of OpenRouter announced the same day.
What It Means for Developers
-
Long context, complex reasoning? K3 may be the best option today — expensive, perhaps, but not necessarily costlier in total.
-
Latency-sensitive? K3 is currently slow — wait for community distillations or smaller variants.
-
Self-hosting? The weights are open, but you need serious compute.
-
Product pricing? K3 proves users will pay a genuine capability premium — provided you can clearly articulate why it costs more.
Model competition is entering the capability-tier phase. Every model finds its position; every task finds its best-suited model — that is what a maturing AI industry looks like. K3’s lesson: parameters can be stacked, but what actually holds value is the ability to innovate on architecture, the authority to set prices, and the leverage of an open ecosystem.
