After Z.AI announced GLM-5.3-Flash midweek, the wave of new models from Chinese labs continued with Tencent and Alibaba just a few days later. Both are pursuing the same direction: open-sourcing, emphasizing low costs and performance on real-world tasks rather than merely chasing benchmark rankings.

Tencent open-sources Hy4 preview: 770B parameters, 1 million token context

On August 28, Tencent released and open-sourced Hy4 preview, the next-generation model in its Hunyuan lineup, featuring 770 billion total parameters, 49 billion active per inference pass, and a 1 million context token window. This is more than triple its predecessor Hy3 (295B total, 21B active, 256K context).

Tencent did not focus on rankings. Instead, the company emphasized performance on practical productivity tasks: long-context coding, office work, financial analysis, game development (generating playable prototypes from natural language prompts), and scientific research. In an internal evaluation involving 163 experts across 203 technical tasks, Hy4 preview scored 2.99/4.00, edging out GLM-5.3 (2.92) and Kimi K3 (2.94).

Most notably, Hy4 preview participated in optimizing its own development process for the first time, proposing training methods, running experiments, and identifying and resolving inference system bottlenecks, which boosted end-to-end throughput by 31.8%. The API is priced at $0.834 per million input tokens and $2.501 per million output tokens. The model is available on WorkBuddy, CodeBuddy, Yuanbao, and OpenRouter.

Alibaba launches Qwen3.8-Flash-Next: 125B total, only 6B active

2 days earlier, Alibaba Qwen announced Qwen3.8-Flash-Next, a multimodal Mixture-of-Experts model with 125 billion total parameters but only 6 billion active per token—nearly three times smaller than Qwen3.7-Plus (17B), while training costs were just 1/9. The model is positioned as an architectural preview for Qwen4, similar to the role Qwen3-Next played for Qwen3.5.

The architecture features four key innovations: hybrid attention (Gated DeltaNet for 3/4 of the layers, Qwen Sparse Attention for the remaining layers), Gated Residual, N-gram Embedding (a table of 51 billion parameters storing common phrases, running on system RAM instead of GPUs), and the Muon optimizer. The native context is 262,144 tokens, expandable up to 1 million using YaRN.

In terms of performance, Qwen stated that Flash-Next outperforms Qwen3.7-Plus on most tasks despite being much cheaper, leading multiple agentic coding benchmarks: 58.7 on DeepSWE, 62.5 on SWE-bench Pro, and 73.9 on CoWorkBench (compared to 45.1 for DeepSeek-V4-Flash), with a JobBench score of 55.7 (nearly double Qwen3.7-Plus). Its weakest area is ultra-hard scientific reasoning (35.9 on HLE compared to 40.0 for Claude Opus 4.6).

Pricing is where it exerts major pressure: $0.16 per million input tokens, $0.47 per million output tokens—roughly 12 times lower than Alibaba’s own flagship Qwen3.8-Max. The production version is Qwen3.8-Flash, which is open-weight on Hugging Face and ModelScope.

What the three companies have in common: competing on cost, not rankings

Three Chinese labs (Z.AI, Tencent, Alibaba) within the same week released open models carrying a similar message: frontier intelligence does not necessarily come at a frontier price. This move signals that Chinese models are entering a new race: instead of proving who is smarter, Chinese companies are competing to prove who is cheaper and more open. Price pressure on Anthropic and OpenAI is real, and OpenAI’s recent price cuts for the GPT-5.6 series may be the first reaction.