Shanghai Dropped a 2.8T Model on Hong Kong

Moonshot AI dropped Kimi K3 — a 2.8-trillion-parameter open-weight model rivaling Anthropic's Fable on coding benchmarks — late on July 16, and on July 17 Chinese President Xi Jinping delivered his first-ever World AI Conference keynote preaching open-source cooperation, wiping as much as 28% off rival Chinese AI stocks in Hong Kong and proving Beijing now treats downloadable weights as diplomacy, not just research.

What landed overnight:

• Moonshot's Kimi K3 tech blog unveiled a 2.8-trillion-parameter sparse MoE model — the first open model to approach the 3T-class — with a 1-million-token context window and native vision.
• Full weights ship by July 27, 2026. API access is live now at $0.30/$3.00/$15.00 per million tokens (cache-hit input / cache-miss input / output).
• Reuters reported the launch came roughly a month after Washington forced Anthropic to withdraw Fable and Mythos over security concerns — reopening the open-weights gap U.S. export controls were meant to close.
• Hong Kong punished domestic rivals Friday: Zhipu fell 27.7% and Minimax 16.5% just before the close, per Reuters.

Xi's stage set:

• Xi delivered his first WAIC keynote on July 17, titled "Joining Hands to Build a Just and Equitable System for Global AI Governance."
• "AI development should not be a solo performance by a single country, but a symphony of international cooperation," Xi said, per Yahoo News.
• He pledged to "encourag[e] open source, openness, collaboration and sharing," SCMP reported — the morning after Kimi K3's debut.
• On July 16, 29 nations signed an agreement to establish the World AI Cooperation Organization in Shanghai, including Russia, Brazil, and Venezuela, per Yahoo.
• Xi's speech landed the same morning Donald Trump accused Beijing of orchestrating the "largest compromise of election data in history" in a prime-time broadcast — dueling narratives on the same news cycle.

What Kimi K3 actually claims:

• Moonshot acknowledges K3's "overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol" — but says it hit frontier-level results on its evaluation suite.
• Where K3 leads: SWE Marathon (42.0 vs Fable's 35.0), Program Bench (77.8 vs 76.8), BrowseComp (91.2 vs 88.0), Terminal Bench 2.1 (88.3 vs 84.6), per Moonshot's benchmark table.
• Where it trails: FrontierSWE (81.2 vs Fable's 86.6), HLE-Full (43.5 vs 53.3), DeepSWE (67.5 vs 70.0).
• On GPU kernel optimization, Moonshot says K3 "performed competitively with Fable 5 (with fallback) and substantially outperformed" Opus 4.8, GPT 5.6 Sol, and GPT 5.5.
• Third-party rankings cited by Reuters: Arena.ai ranked K3 first on web-interface building; Vals AI placed it second overall behind Fable 5 and ahead of GPT 5.6 Sol.

Why the scale matters — and doesn't:

• K3 activates 16 of 896 experts per token via Stable LatentMoE — parameter count is a scale signal, not a capability guarantee.
• Lian Jye Su, chief analyst at Omdia, told Reuters Chinese models gain traction because they "can be run at a fraction of the cost that OpenAI charges its clients" — but cautioned that 2.8 trillion parameters "doesn't necessarily mean you have the best performance by default."
• Ryan Fedasiuk, fellow at the American Enterprise Institute, noted on LinkedIn that running a 2.8T model locally would require hundreds of thousands of dollars in hardware — most users will rent API access, not host it in-house.
• Moonshot recommends deploying on supernode configs with 64+ accelerators for inference efficiency.

The competitive wreckage:

• Kimi K3 is 75% larger than DeepSeek's V4 Pro (~1.6T) and nearly 4× Zhipu's GLM-5 series (~744B), per SCMP.
• Zhipu launched GLM-5.2 to global acclaim just one month ago. Friday's selloff suggests investors now price open-weight leapfrogs as zero-sum inside China's AI sector.
• Moonshot is backed by Alibaba and Tencent and was reportedly seeking $2 billion at a $30 billion valuation ahead of a Hong Kong listing, per Bloomberg reporting cited by Reuters.

What to watch:

• July 27 weight release — the diplomatic promise becomes a downloadable artifact anyone can fork.
• Whether Washington responds with tighter export controls on inference chips or open-weight training data.
• U.S.-China government-level AI talks under the Trump administration — Xi's "symphony" vs. America's gated-frontier strategy.
• Whether Hong Kong's AI stock carnage spreads to U.S. names already selling off on capex anxiety.

Convina's view: Kimi K3 is not proof China has passed American frontier labs — Moonshot's own benchmarks show gaps on the hardest reasoning tasks. It is proof that open weights have become Beijing's answer to Washington's kill switch. Xi preached sharing on a global stage the morning after Moonshot shipped a 2.8-trillion-parameter receipt; Hong Kong sold off everyone who wasn't Moonshot. Enterprises outside China now face a procurement fork: rent gated U.S. APIs subject to export-control whiplash, or build on open weights advancing faster than Western analysts assumed. The six-month lag narrative is dead. The sovereignty question isn't.