On August 13, 2026, Alibaba made its most capable AI model available to the world completely for free.
The model is named Qwen3.8-2.4T-A95B, belonging to Alibaba’s Qwen series. Possessing 2.4 trillion parameters (the adjustable “knobs” within a neural network), it stands among the largest open-source models ever released. The weight files are publicly hosted on Hugging Face, accessible to anyone for download under a license that permits commercial use.
On the exact same day, DeepSeek launched V4 Pro 0813, and Elon Musk’s xAI released Grok 4.6. Three major tech companies releasing major models on a single day would have been unimaginable three years ago. One developer remarked in the comments that the pace of AI model releases has officially entered “China Time.”
What Does 2.4 Trillion Parameters Really Mean?
The human brain contains roughly 86 billion neurons. At 2.4 trillion parameters, this model boasts 28 times that number. Numerically, it resembles packing the neural cell count of 28 human brains into a single entity.
However, there is an essential architectural twist: it does not activate all 2.4 trillion parameters for every response.
Qwen3.8 utilizes a Mixture of Experts (MoE) architecture. Internally, the model maintains 512 “experts.” When presented with a prompt, it awakens only 10 routed experts and 1 shared expert—meaning only about 95 billion parameters actually participate in the active computation. To draw an analogy: imagine a corporation with 24,000 employees, but for any given assignment, only 95 workers report for duty.
The engineering implications are substantial: knowledge is stored across the full 2.4 trillion parameters during training, while inference consumes only about 1/25th of the compute. This is how “massive scale” and “practical affordability” coexist. Compared to traditional dense models of equivalent capability, inference costs are typically reduced by 70% to 80%.
The file size further illustrates the scale: the weight files total approximately 5TB (~5,000GB), enough to fill the hard drives of several personal computers. While certainly not a toy designed for consumer smartphones, its operating costs have been pressed into a viable range by its MoE architecture.
Furthermore, it natively retains a context window of 260,000 tokens—roughly equivalent to a full-length novel—with an extended context capability reaching up to 1 million tokens. From an engineering standpoint, feeding an entire book into a model and asking questions about it was a capability reserved exclusively for closed-source flagships just two years ago.
Why Give It Away for Free?
To clarify the concept of “open source” in this context: all model files are publicly available for any company or individual to download, inspect, fine-tune, or commercialize. By contrast, proprietary models like OpenAI’s GPT series or Google’s Gemini lock their strongest versions behind paywalls, charging per token.
Training a 2.4-trillion-parameter model incurs compute bills in the hundreds of millions of US dollars. Giving such an asset away for free is rare in corporate history.
What is Alibaba pursuing? Three main strategic objectives stand out:
First, ecosystem capture. The base model serves as the entry point. The more developers build on Qwen, the larger Alibaba’s cloud infrastructure and API revenue become. Alongside this open-source release, Alibaba simultaneously announced Qwen3.8-Max—a paid API offering multimodal vision capabilities and a default 1-million-token context window. The free open-source version drives adoption, while the paid version generates enterprise revenue.
Second, competitive pressure. DeepSeek has captured substantial market share through an “open source plus low pricing” strategy, compelling Alibaba to respond aggressively. DeepSeek V4 Pro 0813, released on the same day, priced its input tokens as low as $0.435 per million tokens. In a price war of this intensity, failing to open-source risks falling behind.
Third, regulatory and geopolitical policy. Overseas developers have noted that Chinese authorities explicitly encourage international cooperation in open-source AI, aligning with Qwen’s strategic pivot toward openness after a period of tighter controls.
Engineering judgment: Open-sourcing a flagship model effectively abandons monetization of the model weights themselves in order to capture value across the broader cloud ecosystem. Both business models involve trade-offs, but Alibaba’s strategic bet is unmistakable.
Open Source vs. Closed Source: Which Path Has More Staying Power?
This release highlights the central debate defining current AI development.
The closed-source camp (represented by OpenAI, Google, and Anthropic) argues that as models grow more capable, safety risk increases, necessitating closed iteration to mitigate misuse. Monetizing models directly provides the immense capital needed to fund next-generation R&D. The logic of safety and commercial sustainability is self-consistent.
The open-source camp contends that transparency is essential for auditability, community collaboration accelerates technical progress, and open weights prevent vendor lock-in. As one developer noted on Hacker News, open source empowers individuals to run state-of-the-art AI locally—something closed-source vendors inherently cannot offer.
Both arguments hold merit. However, several empirical trends are undeniable.
Which model will prove more sustainable in the long run remains an open question. Yet one trend is clear: closed source relies on technical moats to sustain premium pricing, whereas open source relies on scale to amortize costs. How long closed-source moats endure depends on how quickly open-source models catch up. In 2026, that convergence is happening much faster than expected.
By 2026, the performance gap between top open-source models and closed-source flagships has narrowed from a generational divide to a few percentage points. Qwen3.8 competes head-to-head with GPT-5.6 and Claude Opus 4.8 across multiple programming benchmarks. On PaperBench, Qwen3.8 scored 93.0, surpassing GPT-5.6 Sol’s score of 90.5.
Skepticism remains. In Hacker News discussions, some commentators described the model card as “too good to be true.” Others pointed out that the release in bf16 precision without native quantization leaves post-compression performance uncertain. Hardware constraints were also highlighted: running even a 1-bit quantized version requires roughly 397GB of VRAM, keeping it out of reach for individual consumer setups.
These critiques are valid. Moreover, parameter races have no fixed finish line—Kimi K3 previously open-sourced a 2.8-trillion-parameter model, and records will continue to be broken. Parameter counts are not everything, and benchmarks do not equal real-world user experience.
Yet one fact is difficult to dispute: in 2026, the open-source ecosystem is capable of competing directly with closed-source giants. Two years ago, few would have anticipated this pace of parity.
Why Should Everyday Users Care?
First, pricing. Open-source models continuously depress market-wide price anchors, making AI products progressively cheaper for everyone.
Second, digital sovereignty and access. This is a tier-one global model developed in China that belongs directly to the user—downloading and commercializing it requires permission from no one.
Third, future integration. The voice assistants on your phone, customer support bots in enterprises, and decision-support systems in healthcare will increasingly run on foundation models made freely available today. What is free today becomes the infrastructure of countless products tomorrow.
The 2.4-trillion-parameter record will eventually be surpassed. However, the decision to make a top-tier model freely available signals a shift in AI competition: from competing purely on parameter scale to competing on who can deliver capability to the widest audience. On that path, open source is leading the way.
A quieter signal remains: when state-of-the-art models become public resources, user access to powerful AI is no longer a corporate favor. It may be early to declare victory, but the direction is undeniable.
Reference Links:
- Hugging Face: Qwen/Qwen3.8-2.4T-A95B
- HN Discussion (HN)
- NVIDIA Technical Blog: Serve Qwen3.8-2.4T-A95B on GB300 NVL72
- Qwen Official X: Qwen3.8-Max Announcement Teaser
- vLLM Recipe: Qwen3.8-2.4T-A95B
Image Verification Note: The original content (Hugging Face model card) consists of plain text and tables without embedded images. All img elements extracted via browser scripts were site logos, avatars, or badges, and were filtered per rules. Therefore, no images are included. Full list:
1. huggingface.co/front/assets/huggingface_logo-noborder.svg (95×88, site logo, filtered)
2. cdn-avatars.huggingface.co/v1/production/uploads/6215ca5692c0ecfba9186921/hrRM50-6XcdWgg2AKpENG.jpeg (200×200, Qwen avatar, 2 occurrences, filtered)
3. img.shields.io/badge/Qwen%20Studio-536af5?logo=qwen&logoColor=white (99×20, Qwen Studio badge, filtered)