Alibaba Announces Qwen3.8: A 2.4T Parameter Open-Weight Powerhouse

Alibaba has announced Qwen3.8, a 2.4T parameter model that will soon be released as open-weight. It is currently ranked as a top-tier model, second only to Fable 5 in performance. Users can immediately test the preview version through Alibaba's Token Plan and Qoder platforms.

Key Points

• Alibaba is launching Qwen3.8, a 2.4 trillion parameter model that will eventually be open-weight.
• The model is claimed to be one of the most powerful available, ranking second only to Fable 5.
• A preview version, Qwen3.8-Max-Preview, is immediately available for testing on Alibaba's platforms.
• The community is showing high engagement, with specific requests for smaller versions to facilitate local deployment.

Sentiment

The sentiment is cautiously optimistic regarding the technology's availability but highly skeptical and analytical concerning the geopolitical motivations and potential for state-mandated censorship.

In Agreement

• Open-weight releases are a strategic move to commoditize intelligence and erode the competitive advantage of closed-source American frontier labs.
• The competition between Chinese firms like Alibaba and Moonshot AI is accelerating the availability of high-quality models for the global community.
• Releasing open weights is an effective soft-power play that builds global goodwill and bypasses geopolitical restrictions on API access.
• Smaller models still have significant headroom for improvement through better post-training and architectural innovations like MoE and 1-bit quantization.

Opposed

• Chinese models may contain 'ideological poisoning' or state-mandated censorship that makes them unsuitable for sensitive or general-purpose use.
• Open-weight models could be 'decelerationist' by destroying the profit margins and R&D budgets required for the next leap in AI progress.
• The environmental and resource cost of multiple firms training nearly identical massive models is a net loss for humanity.
• Open weights are not a substitute for true open source, as the training data and methodologies remain proprietary and opaque.