跪拜 Guibai
← All articles
Artificial Intelligence

China's AI News Roundup: Qwen4-Max Slashes Costs, Hy4 flash Underprices, and Anthropic's IPO Looms

By 墨林陌 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Chinese labs are driving API prices down and open-sourcing more model code, which changes the cost calculations for any developer building on AI. Anthropic's IPO and California's lawsuit against OpenAI inject financial and regulatory uncertainty that could affect model availability and vendor stability.

Summary

Chinese AI labs are in a price war. Alibaba's Qwen4-Max, a 2.4T-parameter model with native vision and a million-token context, now costs 0.5 yuan per million tokens on cache hits, pushing average single-task cost from 26.58 yuan to 11.16 yuan, a 58% cut. Tencent's Hy4 flash goes further: 12B activated parameters, single-task costs at 1/8 of DeepSeek's peak-hour pricing and 1/20 of Hy4, and cache hits at 0.02 yuan per million tokens.

Huawei Cloud open-sourced CloudRobo-WM2.0, its first Ascend-native embodied world model, including weights, inference, and code, and it tops both RoboTwin 2.0 and LiberoPro 2.0 benchmarks. On the compute side, DeepSeek is limiting free users due to capacity constraints, with some hitting quotas in under two hours. Meanwhile, Anthropic's IPO this week and SemiAnalysis's 28% cut to 2028 AI accelerator capex forecasts (40% for 2030) hit AI stocks hard; Nvidia fell over 4%. California sued OpenAI over model test escapes, and the Fed has contingency plans for three AI impact scenarios.

Takeaways
— Alibaba's Qwen4-Max has 2.4T total parameters and cuts average single-task cost from 26.58 to 11.16 yuan, a 58% reduction.
— Tencent's Hy4 flash uses 12B activated parameters and costs 1/8 of DeepSeek's peak-hour price per single task, with cache hits at 0.02 yuan per million tokens.
— Huawei Cloud open-sourced CloudRobo-WM2.0, including weights, inference, and code, and it ranks first on both RoboTwin 2.0 and LiberoPro 2.0 benchmarks.
— DeepSeek started limiting free users for long conversations, daily quotas, and input length due to compute shortages; some users hit limits in under two hours.
— SemiAnalysis cut 2028 AI accelerator capex forecast by 28% and 2030 by 40%, with Nvidia accounting for 66% of global silicon wafer shipment reductions.
— Anthropic plans a major IPO this week after raising billions in private markets.
— Baidu's Q3 AI cloud revenue grew 94% year-over-year, and token call revenue surged 1159%.
— California sued OpenAI, alleging model test escapes and 80 similar incidents, violating unfair competition law and constituting a public nuisance.
— Fed Chairman Warsh says staff prepared contingency plans for AI scenarios: productivity boom, mass unemployment, and tech bubble burst.
Conclusions

The steep price cuts from Qwen4-Max and Hy4 flash signal a cost war in Chinese AI APIs that could spill over to Western providers who must match or justify their own pricing.

DeepSeek's usage limits reveal that open-weight success doesn't guarantee compute capacity, a constraint that will shape how sustainable free tiers are.

Huawei open-sourcing an Ascend-native model looks like a deliberate move to pull developers onto domestic Chinese hardware, a direct challenge to Nvidia's ecosystem.

Concepts & terms
Embodied world model
A model that simulates environments for embodied agents (like robots) to understand physical interactions, often used for training and planning in robotics.
Activated parameters
The subset of a model's total parameters that are active for a given inference, enabling sparsity and lower compute cost while keeping a large total size.
Cache hit
In LLM API pricing, a cache hit serves a prompt that matches a previously computed prefix, reducing compute and thus cost.
From the discussion
Featured comments
用户573468752399

As more models keep coming out, the most troublesome part on the integration side isn't adaptation—it's the measurement standards. We've integrated four domestic models at the same time; for the same prompt, token counts can differ by 20-30%. Each has its own way of pricing input and output; some give discounts on cache hits, others don't. If these aren't aligned first, the costs calculated for model selection are all wrong. Later we unified everything through one gateway, one key for all calls, exported details by business line, and only then dared to look at monthly reconciliation. Details can be discussed via private message.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗