Alibaba launches Qwen3.8-Max with 2.4 trillion parameters

Alibaba has entered the trillion-parameter arms race with a 2.4 trillion-parameter model it claims trails only Anthropic's Claude Fable 5.

Alibaba released Qwen3.8-Max-Preview, a 2.4 trillion-parameter AI model that the company says is the second-most capable system available globally, behind only Anthropic's Claude Fable 5. The model launched on Alibaba's new Token Plan Personal Edition, with the official version and open weights expected soon.

"Qwen3.8-Max demonstrates strong performance in code engineering and professional office scenarios, making it one of the most capable models available today," Alibaba said in its announcement, positioning the system as a direct competitor to frontier models from Anthropic, OpenAI and Google.

The model arrives less than three months after Qwen3.6 Max Preview, which launched in April 2026 as Alibaba's first trillion-parameter release. At 2.4 trillion parameters, Qwen3.8 enters the same scale bracket as Moonshot AI's Kimi K3 at 2.8 trillion parameters — a notable dynamic given Alibaba owns roughly 36% of Moonshot, the Beijing-based lab behind Kimi. Alibaba's first trillion-parameter model, Qwen3-Max-Preview, debuted in September 2025.

The launch deepens an accelerating trend of Chinese AI labs challenging Western dominance. Three major Chinese releases in as many months — Z.AI's GLM 5.2, Moonshot's Kimi K3, and now Qwen3.8 — have each claimed positions near the top of global leaderboards. Kimi K3 briefly held the title of the largest open-source model ever released and beat Claude Fable 5 and GPT-5.6 Sol outright on Arena's Frontend Code leaderboard, becoming the first open-weight model to top that ranking. GLM 5.2 landed within striking distance of Claude Opus 4.8 on FrontierSWE.

Alibaba has not yet published independent third-party benchmarks for Qwen3.8, meaning the "second only to Fable 5" claim currently rests on internal evaluations — the same approach Moonshot took when first announcing Kimi K3. Whether Qwen3.8 holds up once outlets like Artificial Analysis and LMArena run their own tests will determine how much of that claim survives independent scrutiny.

The Token Plan Personal Edition pricing starts at $6 per month for the Lite tier, offering 2,500 credits every seven days, and goes up to $68 per month for Pro, which provides 40,000 credits over the same window and supports running six to eight agents concurrently. The plans carry compatibility with tools built around OpenAI and Anthropic's protocols, allowing developers using existing coding agents to point them at Qwen3.8 without rebuilding workflows. Team editions range from $150 per seat per month to $1,398 per seat per month.

For investors, the question is whether Alibaba can convert model capability into cloud revenue. Alibaba Cloud has been the company's fastest-growing segment, and the Qwen3.8 launch — bundled into subscription plans — signals a push to monetize its AI investment directly. The broader competitive picture also pressures Nvidia, whose H100 GPUs power much of the training infrastructure for these models, though Alibaba has been diversifying its chip procurement with in-house and third-party alternatives. If Qwen3.8's performance claims hold, it could shift enterprise AI spending toward Alibaba's ecosystem and away from Western API providers, particularly for cost-sensitive workloads in Asia.

This article is for informational purposes only and does not constitute investment advice.