Key Takeaways:
Tencent's Hy4 giveaway is the fourth frontier-model release from Chinese labs in a month — a pricing strategy, not a release calendar.
Tencent open-sourced its 770 billion-parameter Hy4 preview under an Apache 2.0 license on Friday, the fourth frontier-model giveaway from Chinese labs in a month, a coordinated push to price AI intelligence toward zero and squeeze Western closed-API rivals.
"A free frontier model works on the market the way a price signal does. Every closed lab now has to justify its margin against a competitor charging nothing for the weights," Jon Markman, a Forbes contributor covering technology stocks, wrote in his analysis of the release.
Hy4 activates 49 billion parameters per token and reads a million tokens of context, with hosted access at about 83 cents per million input tokens. Simon Willison's early review treats it as a serious coding and agent model aimed at the productivity work the paid frontier tier charges for. The release follows DeepSeek's MIT-licensed V4-Flash (284 billion parameters, 14 cents per million input tokens) on July 31, DeepSeek's V4-Flash-Vision-Exp on August 21, and Alibaba's QwenWork agent platform opening to international users on August 26.
The strategy targets a market where Chinese labs face trust and procurement barriers selling closed AI services into Western enterprises, whatever their benchmarks show. By giving away weights, they commoditize the product their competitors sell — and shift pricing power toward the compute layer beneath it, where China's top four cloud platforms are set to spend 726 billion yuan on capital expenditure in 2026, up 98 percent year over year, according to BofA Securities.
The pricing gap the strategy exploits is stark. Anthropic's Claude Opus 5 charges $5 per million input tokens and $25 per million output tokens, while OpenAI's GPT-5.6 Sol runs $4 and $20, per Artificial Analysis data cited by BofA. Among Chinese models, Moonshot AI's Kimi K3 prices at $3 and $15, Alibaba's Qwen-3.8 Max at $2 and $6, and DeepSeek V4.0 at just 70 cents and $2. On a per-task basis, Claude Fable 5 costs $3.14 per intelligence-index task versus 84 cents for Kimi K3 and 68 cents for Zhipu's GLM-5.3.
The capability gap has narrowed to the point where the price difference is hard to justify. Kimi K3 scores 60 on Artificial Analysis' intelligence index, about 95 percent of Claude Opus 5's 63, with GLM-5.3 tied at 60. On the front-end web development benchmark Code Arena WebDev, Claude Opus 5 leads at 1,691 points, with Kimi K3 at 1,674 and Qwen-3.8 Max at 1,669 — Tencent's Hy4 preview ranks sixth at 1,633.
The weights already travel even where the brand cannot. Cursor's Composer model is built on Moonshot's Kimi K2.5 checkpoint, a Chinese open-weight base that an American company post-trained into near-frontier coding performance. DeepSeek V4 Flash led all models with 22 percent of weekly token usage on OpenRouter as of mid-August, up 4.7 percentage points from July, while Anthropic still commands 64.8 percent of spending share on the Vercel AI platform — a gap that reflects the pricing differential more than capability.
The squeeze lands on whoever sells intelligence at a markup without owning distribution. Mid-priced closed APIs now compete with Apache-licensed systems at a tenth of their price, and application wrappers whose main asset is a system prompt on someone else's model watch their input cost and differentiation fall together. Frontier labs keep their edge only at the top of the capability range, and that band narrows with every strong open release.
Compute is the layer the strategy cannot touch. Free weights still need GPUs to run on, power to feed them, and racks to hold them — and cheaper intelligence has so far expanded total usage rather than shrinking the spend beneath it. BofA expects China's cloud market to grow from 389 billion yuan in 2025 to 1.73 trillion yuan by 2030, a 38 percent compound annual rate, with AI-related cloud services rising from 17 percent to 66 percent of the market. The AI model-as-a-service segment alone is projected to grow at an 81 percent compound annual rate to 676 billion yuan by 2030.
Tencent's cloud revenue grew about 20 percent year over year in the second quarter, while Alibaba's rose 45 percent, and consensus estimates for both companies' next-12-month capex have been raised 25 percent and 29 percent respectively over the past 30 days, per BofA. The confirming evidence from here is commercial: whether Western closed labs cut API prices this autumn with Hy4 sitting on public leaderboards, whether more Western products ship on Chinese checkpoints the way Cursor did, and whether enterprise pilots land once procurement teams compare an Apache license against a per-seat contract. The model layer is being priced like a commodity by the companies least able to sell it any other way — and what cannot be given away, the chips, the power, and the data centers, is where the pricing power went.
This article is for informational purposes only and does not constitute investment advice.
