Alibaba Qwen Release Puts Open-Model Pressure Back On US AI Labs

Chinese AI developers are turning open weights into a pricing and deployment challenge for US model labs.

The Register reported that Alibaba launched Qwen 3.8-Max, a 2.4 trillion-parameter model, while DeepSeek's V4 Flash refresh arrived with a smaller footprint and lower task cost.

The movement is not only a benchmark contest.

For enterprise buyers, the practical question is whether capable models can be downloaded, fine-tuned and run on owned infrastructure instead of remaining tied to proprietary cloud APIs.

The Register framed Alibaba, DeepSeek, Moonshot, MiniMax and Z.ai as the Chinese developers pushing hardest into that opening.

Alibaba Sets Qwen Weight Release

Alibaba had kept its strongest models behind an API.

Qwen 3.8-Max changes that by putting the company's most capable model weights on a release path for popular repositories, including Hugging Face, starting next week.

The model is large enough to keep deployment difficult.

The report said Qwen 3.8-Max has 2.4 trillion total parameters and uses 95 billion active parameters for a request.

Its hardware estimate puts customer-facing deployment at 48 to 64 Nvidia B200-class GPUs, while internal workloads would still need 8 to 16 B300 or AMD MI355X GPUs.

Alibaba also plans a 27-billion-parameter version alongside the Max model.

A smaller weight release gives enterprise teams a more realistic path for budgeting, hardware planning and fine-tuning workflows.

DeepSeek Prices V4 Flash

DeepSeek's V4 Flash refresh attacks the same market from the opposite direction.

DeepSeek V4-Flash-0731 is a 284-billion-parameter model that The report said can fit into about 142 GB of GPU memory at FP4, making large-scale local operation possible on a single system.

Artificial Analysis benchmarks, as cited by The Register, put DeepSeek V4 Flash close to OpenAI's GPT 5.6 Luna and assigned DeepSeek a 40 percent task-cost advantage.

DeepSeek API pricing was listed at $0.14 for each million input tokens and $0.28 for each million output tokens; Alibaba's QwenCloud listing for Qwen 3.8-Max was $2 for each million input tokens and $6 for each million output tokens.

The cost comparison is not limited to posted token prices.

Reasoning and agent workloads can consume different token volumes, so model efficiency matters when developers run large coding or automation jobs.

DeepSeek's integration of DSpark speculative decoding is presented as one reason for the efficiency claim, with DeepSeek saying it can deliver 57 to 85 percent more per-user speed on the same hardware.

Buyers Compare Closed APIs With Weights

US and European AI companies have warned about Chinese open models, safety standards and model provenance.

Hugging Face CEO Clément Delangue said on CNBC that Chinese developers are clearly dominating open models and could start dominating frontier models by the end of this year or next year if the pace continues.

That leaves US labs with a narrower response than broad safety rhetoric.

Enterprise customers weighing proprietary APIs against open models will compare control, auditability, data handling, price and available hardware.

If Chinese models keep closing benchmark gaps while offering downloadable weights, the competitive question becomes how much extra trust or capability customers get from closed systems.

Production use will decide how much of the launch economics survives outside charts.

Alibaba's highest-end model still requires costly infrastructure, and DeepSeek's lower-cost path depends on whether compressed or quantized deployments preserve output quality under real workloads.