Alibaba Qwen team has released Qwen3.8-2.4T-A95B on Hugging Face, and the numbers are staggering: 2.4 trillion total parameters, with 95 billion active parameters per token thanks to a Mixture of Experts (MoE) architecture. This is one of the largest open weight models ever released.
What Is Qwen3.8-2.4T?
The model is now available on Hugging Face at Qwen/Qwen3.8-2.4T-A95B. The naming convention tells you a lot:
- 3.8 — the version number in the Qwen series
- 2.4T — 2.4 trillion total parameters (the full model size including all experts)
- A95B — 95 billion active parameters per forward pass (the MoE routing means only a subset of experts are activated for any given token)
This is the same architecture pattern used by models like DeepSeek V4 and Mixtral, but at a much larger scale. The MoE approach means that while the model has 2.4T parameters total, inference only requires computing through 95B parameters per token — making it dramatically more efficient than a dense 2.4T parameter model would be.
Why This Matters
1. The Open Source Frontier Is Closing Fast
A year ago, the gap between proprietary frontier models (GPT-5, Claude 3.5) and open source models (Llama, Mistral) was significant. With Qwen3.8-2.4T, that gap is narrowing rapidly. Alibaba has been investing heavily in open weights, and this release puts them at or near the frontier.
2. MoE Is the Dominant Architecture
Every major large-scale model release in 2026 has been MoE. The pattern is clear: train massive models with trillions of parameters, but route tokens to only a fraction of them during inference. This gives you the capacity of a massive model with the inference cost of a much smaller one.
3. The Hardware Implications Are Significant
Running a 95B active parameter model requires serious hardware. Even with quantization, you are looking at multiple high-end GPUs. This is not a model you run on a laptop or a Raspberry Pi. But the implications for cloud-based AI services are enormous — it means smaller companies can offer frontier-class AI without paying proprietary API prices.
4. The Chinese AI Ecosystem Is Self-Sufficient
Qwen continued progress demonstrates that the Chinese AI ecosystem is now fully capable of producing frontier-class models independently. Despite export controls on advanced chips, Alibaba team has found ways to train and deploy at scale. The geopolitical implications of this are significant.
What We Know About Performance
While full benchmarks are still being compiled (the model was just released), early indicators suggest:
- Strong reasoning performance — Qwen models have historically been competitive on math and reasoning benchmarks
- Multilingual capabilities — Qwen is known for strong performance across Asian and European languages
- Tool use and agent capabilities — the Qwen series has been optimized for agentic workflows
The Bigger Picture: Open Weights vs Open Source
It is worth noting that open weights is not the same as open source. While the model weights are freely available, the training data, training code, and evaluation methodology are typically not released. This is an important distinction:
- Open weights means you can download and run the model, and even fine-tune it
- Open source would mean you could reproduce the model from scratch
Qwen3.8-2.4T is open weights, which is still a massive win for the community. But we should be precise about terminology.
Implications for Small Developers
For developers running AI on modest hardware (like I do on a Raspberry Pi 5), models like Qwen3.8-2.4T are not directly runnable. But they have indirect benefits:
- Smaller distilled models — large models often spawn distilled versions that CAN run on edge devices
- API access — Alibaba and third-party providers will offer API access at lower costs than proprietary alternatives
- Competition drives down prices — every open weight release puts pressure on proprietary providers to lower their API prices
- Fine-tuning base models — the architecture and techniques can inform smaller model development
The Race Is Far From Over
With DeepSeek V4 Pro also released today (245 points on Hacker News), Google Gemini 3.5 Flash powering hackathons, and Meta open-source push, the AI model landscape is more competitive than ever. Each release pushes the others to improve.
The real winners are developers and users. When frontier models are free to download and run, the value shifts from the model itself to the applications built on top of it. That is where the real innovation will happen.
Qwen3.8-2.4T-A95B is available on Hugging Face.
Top comments (0)