“Open weights” no longer means one thing: three business models share the label — resolves 2027-07-28

The debate: Chinese model makers have been publishing their model weights (the trained parameter files that let anyone run the model themselves), and outsiders lump them all into one “open source / open weights” camp.

Why it matters: a buyer who picks vendors by that label can misjudge what license it actually gets and how much it will pay.

Who is on each side: the price-crash camp says open weights will flatten prices across the whole model layer; we took the label apart and found three completely different business designs.

“Open weights” no longer means one thing: it covers licensing thresholds (Moonshot, MiniMax), withheld flagships (Alibaba, gpt-oss), and fully open releases monetized elsewhere (Zhipu, DeepSeek). China's price compression survives industry-wide only where a near-frontier vendor stays in the cheap tier. K3's license is documented; its top-tier price is not readable.

Moonshot AI (via HuggingFace)
“If the Licensee or any of its affiliates operates…”

ChinAI Newsletter (Jeff Ding, GWU)
“Using the blended rate (essentially, a typical…”

X / @natolambert (Nathan Lambert, Ai2/Interconnects)
“Kimi K3 license. It's inspired by MIT but…”

llm-stats.com + DevTk.AI + pricepertoken.com + Yotta Labs (independent price trackers, unaffiliated with Moonshot, ChinAI or Lambert)
“GLM-5.2 (via DeepInfra): Input $0.95/1M tokens…”

BG2 Pod (Brad Gerstner + Bill Gurley)
“in the in the last one in the 14th they literally…”

Moonshot AI (official pricing page)
“¥2.00 / MTok”

DeepSeek (official pricing page)
“1M INPUT TOKENS (CACHE MISS) $0.435”

Five things to watch: (1) K3's effective price falling back into DeepSeek's band, blended under $0.50, before mid-2027, or Moonshot cutting its price or loosening its license, voids the monetization reading; (2) Moonshot holding at the top of the band after expanding capacity, with the resumption of subscription sales as the signal that capacity has expanded, weakens the rationing reading; (3) Zhipu or DeepSeek also gating their next flagship reopens the industry-level conclusion about the compression mechanism; (4) public evidence of cloud vendors signing separate agreements, or of certified inference partners, means the threshold is moving into enforcement; (5) the license attached to Qwen3.8 Max at release, which the company so far describes only as coming soon, is the nearest test of whether gating is an isolated case or a turn.

What would overturn it

If K3's licensing threshold is removed before it expires, or Moonshot restores unrestricted licensing, the split reading collapses back into a single open-weights story and the license-as-hard-evidence call is overturned.

Read the full deep dive →

Jul 28, 2027

Near-term checks: the release of Qwen3.8 Max's weights (officially “soon”, no date); Moonshot's Hong Kong listing (reportedly within 6 months, around 2027-01); when K3 subscriptions reopen