On the evening of July 27, Moonshot AI dropped Kimi K3’s complete model weights onto Hugging Face — Modified MIT license, 2.8 trillion parameters. Within 30 minutes: 4,000+ likes and the number-one trending slot. Hugging Face CEO and co-founder Clem Delangue called it the fastest-growing release in the platform’s history.
Less than a month earlier, the same model had crushed Moonshot’s own consumer cluster, forcing a pause on new subscriptions. Then they gave away the world’s largest open weights. The two moves look contradictory. They are two entirely different lines of business.
An Earthquake in Silicon Valley
The chain reaction ran well beyond a model launch. Vercel CEO Guillermo Rauch called K3 “the strongest open-weight model in the world, full stop.” vLLM shipped support the day the weights landed; Fireworks AI, Together AI, and Modal announced Day-0 integrations in sequence.
But what truly detonated in Silicon Valley was not the model — it was the rift it tore open. NVIDIA and Anthropic took opposite sides in public.
On July 22, Jensen Huang told Axios that Chinese open models are “excellent” and US companies “absolutely should be allowed to use them.” He aimed directly at Anthropic: restricting Claude Mythos over cybersecurity risk is the wrong call — powerful models should be tested, the problems found, the fixes shipped. Two days later, 25 institutions including NVIDIA, Microsoft, Meta, Hugging Face, and AMD published the open letter “Open Weights and American AI Leadership”; Google and OpenAI joined after, pushing signatures to 77.
Anthropic was the only US frontier lab absent. Under pressure, CEO Dario Amodei published a long response the same day, insisting “Anthropic has never argued for banning open-weight models.” But read closely, the core positions are unchanged: no high-end chips to China, crack down on industrial-scale distillation, safety-test every strong model. Translated: China’s open-source play is wrecking our business model, and we are losing customers.
NVIDIA then formed an “Open Safe AI Alliance” with Microsoft, IBM, Hugging Face, and SpaceXAI — this time without OpenAI and Google. Everyone is defending the version of “open” that fits their position in the value chain.
Three Hammers From Chinese Open Models
K3 was no isolated event. It was the third strike in three months:
First hammer — DeepSeek V4 preview (April 24). 1.6T parameters, 49B active, million-token context. Performance near Claude Opus at a price that defies comparison: $0.87 per million output tokens, about 1/29 of Claude Opus 5.
Second hammer — GLM-5.2 (June 16). MIT license, million-token context, built for long-haul work — reading large codebases, chaining tools across hundreds of turns. Terminal-Bench 2.1: 82.7%, closing on GPT-5.5’s 83.4%.
Third hammer — Kimi K3 (July 27). The first open model at the 3-trillion scale: 2.8T total, 104B active per token, image and video input, million-token context. Terminal-Bench 2.1: 88.3%, past Claude Fable 5’s 88.0% and within 0.5 points of GPT-5.6 Sol’s 88.8%. On SWE-Marathon, K3 scored 42 — above GPT-5.6 Sol’s 39 and Claude Opus 4.8’s 40.
The three hammers hit price, long-task capability, and scale respectively — each one demolishing a closed-model exclusive.
The People Who Can’t Run It Are the Real Buyers
Open K3’s Hugging Face page and there sits a 594GB weight file. Most people quietly close the tab — and for good reason. After MXFP4 quantization the model needs 8× H100 80GB just to load; production deployments typically call for supernodes of 64+ accelerators, a hardware bill starting in the millions of dollars. A single RTX 4090 cannot hold one of K3’s expert modules. AMD validated an 8× MI355X configuration at roughly 1.7–2.4 million RMB for the machine — and that is “loads,” not “serves.” With 896 experts and 16 activated per token, cross-card communication overhead is brutal. Kimi’s official guidance of 64+ cards puts real commercial deployments in the ten-million-RMB range.
And yet the people who cannot run it are the ultimate buyers. On launch day, Tencent Cloud ADP, Alibaba Cloud Bailian, TokenHub, and the national supercomputing portal all opened K3 endpoints. No H100 purchase, no data center — just call the API and use a 2.8T foundation. Small businesses moved from “can’t afford GPUs” to “which cloud should we pick?” — and bargaining power appeared. Previously only Moonshot served K3’s API, with pricing power entirely in the vendor’s hands; with multiple clouds hosting simultaneously, long-term pricing cannot be locked by one player.
The Modified MIT license is the other key: commercial use, modification, and redistribution all permitted — a company can build closed-source industry-tuned models on K3 without exposing its business code. Closed APIs will never grant that right. And MoE’s sparsity hides a distillation dividend: communities can distill just the frequently-activated expert subsets into “K3 student models” that run on one card. Expect 70B and even 7B distillations within months.
The Next Move in the Game
K3’s biggest change: frontier intelligence increasingly behaves like a downloadable, transferable base asset. As model-layer scarcity falls, value does not vanish — it migrates to compute, inference, agent frameworks, enterprise data, application experience, and safety governance.
But the story is unfinished. Moonshot’s own model card concedes that despite benchmark parity, K3’s overall user experience still trails Claude Fable 5 and GPT-5.6 Sol by a noticeable margin. Enterprises pricing models need to count rework time alongside API fees — a long task that fails at the end costs engineer hours that quickly eat the savings. That is the last wall DeepSeek V4’s official release must break: tool calls, long context, and multi-round edits that stay stable across repeated runs, so developers dare hand large codebases and long agent tasks to open models.
Meanwhile, Washington is moving: OpenAI and Anthropic are pushing distillation investigations, discussing sanctions and access controls. But once weights are downloaded and deployed at scale, restrictions lose most of their bite. The race now is simply who arrives first — DeepSeek V4 official, or Washington’s restrictions.
What To Do With It
1. Model choice is widening. Closed vendors used to hold the model in one hand and the price tag in the other, with no seat at the table for you. K3 put modification rights on the table — and even if you can’t run it locally, cloud APIs are ready.
2. Don’t be hypnotized by parameter counts. 2.8T is a headline; deployment economics are the decision. Watch inference cost, cache-hit rates, and production stability instead of the trillion figure.
3. Watch the ecosystem, not the model. K3’s real value will be decided by whether the community distills usable student models in the coming months, whether hosted prices keep falling, and whether the license stays permissive. Any one of those tightening, and small businesses’ options shrink fast.
Open models may not end closed models — but they are forcing the closed ones to cut costs. Either way, for developers and users, that is good news.
References: 36Kr/APPSO, Zimu AI, AI Changfandiao, QbitAI
