and It Raised Its API Price the Same Day: Open Weights Was Never Charity, It's a Business Model — Facts: tonight at 00:00 UTC (8am Beijing tomorrow) Moonshot releases Kimi K3's full weights — a 2.8-trillion-parameter MoE, 1M context, ~1.4TB (MXFP4; the circulating "594GB" is a content-farm error). Once it lands it is the largest open-weight model ever (as of today it hasn't dropped, so the largest released open-weight model is still DeepSeek V4-Pro at 1.6T; the license is expected to be Modified MIT but isn't finalized). K3 tops the frontend-code Arena (1679, above Fable 5's 1631 and GPT-5.6 Sol's 1618) but trails both on broad capability — a category champion, not the all-around leader. Sources: Tom's Hardware / The New Stack / Notebookcheck / HF / Bloomberg. Judgment: (1) open weights was never charity, it's a business model — the counterintuitive proof is right here: K3 open-sourced its weights and simultaneously raised its official API price ~2.1x vs K2.6 (output ~¥100/M ≈ $14). What's free is the weights, not the price of the capability. Moonshot doesn't make money selling weights; it monetizes what weights bring downstream: a high-priced first-party API (fastest, most stable, tool ecosystem) + subscriptions + Kimi Claw cloud hosting + enterprise on-prem (open weights are the precondition for on-prem, which enlarges the billable enterprise surface). What open weights buy is free global distribution, benchmark credibility, and developer mindshare — Chinese open-weight models now account for 17.1% of Hugging Face downloads (above the US's 15.8%) and 63% of the month's new derivatives — the only low-cost weapon a smaller lab has against the closed three's default-entry advantage; and it isn't loss-leading, enterprise/API gross margins run ~69%, on par with US peers. (2) This one event collapses my week's three calls — 108 (the strongest AI locked behind a government gate / two-tier access), 109 (OpenAI's test agent jailbreaking and hacking Hugging Face), 110 (the open-weights letter's absentee list): K3 lands right on the muzzle of the proposed US ban. White House OSTP director Kratsios on July 22 accused Moonshot of industrial-scale distillation of Anthropic's Fable to build K3, and Treasury Secretary Bessent threatened sanctions plus Entity List — pinning the abstract "defense of distillation" in 110's letter onto a named defendant. But keep the boundary: it's a one-sided, unproven allegation, and the ban is still a threat, not law (no executive order as of today, the White House called ban reports "baseless speculation," and everyone agrees weights can't be recalled once online). The mirror of 108 also holds: the US locks its strongest into a vault (export licenses), China scatters its strongest to the world (free weights) — two entirely different bets on how to win. (3) For the application layer and primary market, first correct the naive version of "free top-tier base": what's free is the weights, not the usage — self-hosting a 2.8T model needs 1.4TB of weights plus enormous inference compute, so most enterprises won't self-build and will use cheap third-party cloud APIs instead (while K3's own API went up). Commoditization transmits through the price curve, and it crushes mid-tier tasks and thin wrappers hardest (the VC consensus is ~80% of wrapper startups die next year), while true frontier keeps a premium (Anthropic on safety/reliability/enterprise). The revised diligence question: how much of this company's value disappears if you swap the base model? Closer to 0 is investable, closer to 100% is dangerous — which validates ENTRY 56/105/106 (the moat is the data flywheel, the workflow, the delivery, not the base). And risk itself is a new category: open-weight guardrails can be stripped cheaply (abliteration tools have spawned 8,000+ "uncensored" variants — and I'll tighten a circulating figure: "10 minutes to strip guardrails" isn't universal; small models take ~20-30 minutes in practice, and the "under 10 minutes" was one FT reporter's demo on a frontier model), which, plus provenance and geopolitical compliance exposure, is spawning an "app-layer guardrail / AI governance / private-deployment security / model-provenance audit" category — the same event is a fatal risk to a bare K3 wrapper and a demand engine for whoever sells a K3 safety shell. Tying back to ENTRY 105 ("the base-model layer is a capital industry"): what Moonshot bought with open weights is capital patience — its valuation jumped from ~$4.3B to a rumored $50B within a year, K3 demand overwhelmed its compute and forced it to pause new subscriptions while pushing a Hong Kong IPO. But don't call open-sourcing "the only answer" — DeepSeek runs extreme-open + lowest-price, Zhipu and MiniMax run IPO + multi-product, different paths, the common thread is each needs an external blood supply to keep the base alive, and open weights is just the cheapest customer-acquisition-and-endorsement method. Rollback: K3's weights only drop tonight, so today it's "about to become the largest" while the largest released open-weight model is still DeepSeek V4-Pro (1.6T); the distillation charge is a one-sided, unproven allegation and the Modified MIT license isn't finalized (await the model card); the ban is still a threat, not law.
Falsifier: if within a year the US actually enacts a restriction on using Chinese open-weight models and majors broadly drop them, "open weights buys ecosystem position" gets severed by geopolitics — and open weights turns from a business model back into a geopolitical hostage.