Alibaba's Qwen team announced Qwen3.8, a 2.4 trillion parameter model going open-weight soon. This would dwarf DeepSeek-V3's 671B total parameters and challenge Llama 4's estimated 2T scale.
Key facts
- 2.4 trillion parameters — largest open-weight model announced.
- Dwarfs DeepSeek-V3's 671B total parameters.
- Exceeds Llama 4's estimated 2 trillion parameters.
- No benchmark scores or architecture details disclosed yet.
- 4-bit quantized version would require ~1.2 TB GPU memory.
Alibaba's Qwen team announced on X that Qwen3.8 is launching and going open-weight soon, with a massive 2.4 trillion parameters. This parameter count would make Qwen3.8 the largest open-weight model ever released, dwarfing DeepSeek-V3's 671B total parameters and Llama 4's estimated 2T.
The announcement comes as the open-weight frontier model race has intensified over the past 12 months. DeepSeek-V3, released in December 2024, demonstrated that a Mixture-of-Experts architecture with 671B total parameters could rival GPT-4 on several benchmarks. Meta's Llama 4, reportedly in the 2 trillion parameter range, has not yet been released publicly. Alibaba's Qwen3.8 leapfrogs both in raw scale, but raw parameter count alone does not guarantee superior performance—training data quality, architecture choices, and inference efficiency matter as much.
What the announcement lacks
Alibaba has not disclosed the training compute budget, dataset composition, or architecture details beyond the parameter count. The Qwen team's tweet says the model is "continuously evolving," suggesting it may still be in active training. No benchmark scores, context window length, or inference latency figures were shared. Without these, the claim remains a headline-grabbing parameter count rather than a verified capability.
The open-weight strategy
Open-weight release of a model this large is unprecedented. Previous 1T+ parameter models—like Google's PaLM 2 or OpenAI's GPT-4—were kept proprietary. If Alibaba delivers on the open-weight promise, it would give the research community and startups access to frontier-scale capabilities without API dependency. However, serving a 2.4T parameter model locally would require massive hardware: even at 4-bit quantization, it would demand over 1.2 TB of GPU memory, limiting deployment to high-end clusters.
Competitive implications
Alibaba's move pressures Meta to accelerate Llama 4's release and forces DeepSeek to respond with an even larger model. The Chinese AI ecosystem, already producing competitive open-weight models (Qwen2.5, DeepSeek-V3, Yi-Lightning), is now competing on scale directly. Western labs may face pressure to open-weight their largest models or risk losing developer mindshare.
Key Takeaways
- Alibaba's Qwen3.8, a 2.4T parameter open-weight model, was announced.
- It would be the largest open-weight model ever, but lacks benchmark details.
What to watch
Watch for Alibaba to release technical report details—likely at an upcoming developer conference or alongside the weights—to validate the claimed scale. Also track Meta's response on Llama 4 timeline and DeepSeek's next model announcement.
[Updated 19 Jul via scmp_tech]
Alibaba has made a preview version, Qwen3.8-Max-Preview, immediately available on its Token Plan subscription service and on its Qoder and QoderWork agentic platforms, according to an official X post on Sunday [per SCMP]. The company claims the model is "second only" to Anthropic's Claude Fable 5, directly positioning it against a known frontier competitor rather than just touting parameter count.
