Alibaba unveiled Qwen 3.8 in preview: 2.4 trillion parameters, multimodal (images, video, documents), early access via Token Plan, Qoder, and QoderWork at 10% standard pricing. Open weights arriving “soon.” The Qwen team claims the model is “second only to Claude 3.5 Sonnet” and outperforms Qwen 3.7-Max on coding, full-stack development, data analysis, and office workflows. No published benchmarks.

As we reported on July 17 with Kimi K3 from Moonshot AI, Chinese open-weight models are compressing the gap from frontier-tier at a pace reshaping decisions for everyone. If the gap between the best open model and the best proprietary model measured around a year at the start of 2025, it’s now dropped to six months or less.

Why this matters to you. Unless you face strict privacy or data sovereignty constraints, the question “frontier model or open model?” is becoming a cost choice, not a quality one. The quality leap from Qwen 3.8 and Kimi K3 means that for many daily tasks (from coding to document analysis) an open-weight model at a fraction of the price delivers usable results. Alibaba’s move explicitly targets Kimi K3 momentum: Moonshot hit $300 million ARR in June and aims for IPO in six months, and Qwen wants to capture that demand before it solidifies.

Before trusting “second only to Claude 3.5 Sonnet,” do what we suggest in the playbook Compare two models in 15 minutes: same task, same test cases, results table. Without published benchmarks, your verification beats the press release.

In detail

This move makes sense against recent dynamics. On July 17, Moonshot AI released Kimi K3: 2.8 trillion total parameters, 50 billion active, the largest open-weight model published at the time. Performance was close to Claude 3.5 Sonnet and GPT-4o, with competitive API pricing. The signal was twofold: China closed the quality gap, but raised prices versus the near-free Chinese models era.

Qwen 3.8 answers K3 on two fronts. Technically, it scales to 2.4T parameters (K3 is 2.8T but with MoE architecture, so active parameters differ) and adds multimodality: images, video, documents. Developer Shuai Bai describes it as the first multimodal Qwen model above 1 trillion parameters. Commercially, Alibaba offers preview access at 10% standard pricing, a move clearly designed to intercept users evaluating Kimi K3 or using it through chat apps and APIs.

What “open-weight” means here. The model’s weights (the numbers that make it work) will be published and downloadable, so you can run it on your servers or a cloud of your choice instead of calling Alibaba’s API. The exact license isn’t announced yet. “Open-weight” doesn’t equal “open-source” in the classical sense: training code and data aren’t released, but weights are, enough for self-hosting or marginal-cost inference.

Gaps in what we know. Alibaba didn’t publish benchmarks. The “second only to Claude 3.5 Sonnet” claim is marketing until verifiable numbers exist. The model is in preview, not general release: open weights arrive “soon” with no date. Competition with Kimi K3 is real, but both are from Chinese companies with local regulatory compliance obligations, which for some sectors (healthcare, finance in some countries) remains an exclusion factor regardless of quality.

The concrete implication. If you’re a professional using AI for recurring tasks with no privacy constraints forcing local models or specific data residency, the open-weight landscape now gives you a credible frontier alternative at a fraction of the cost. Frontier models aren’t obsolete: where maximum quality matters (research, mathematical reasoning, high-stakes work), a six-month gap can weigh. It means the choice stopped being obvious and needs to be made on your cases, not the rankings.