Qwen opens a 2.4T model to self-hosting, with 95B active parameters

On August 12th, Alibaba Cloud's Qwen team released Qwen3.8-2.4T-A95B, an open-weight mixture-of-experts model with 2.4 trillion total parameters and 95 billion active during inference, making it the first Qwen-Max-class model with downloadable weights. The release offers a self-hosted alternative to OpenAI's GPT-5.5 Pro, with adjustable reasoning depth and support for Hugging Face Transformers, vLLM, SGLang, and TokenSpeed, though running it requires substantial infrastructure.

Qwen opens a 2.4T model to self-hosting, with 95B active parameters The Alibaba Cloud model activates 95B parameters per token, supports adjustable reasoning depth and offers a self-hosted alternative to OpenAI's GPT-5.5 Pro. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: GitHub Gist https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3 Why it matters Qwen3.8 gives infrastructure teams downloadable weights for an Alibaba Max-class model with 95B active parameters, creating a self-hosted option for organizations that prioritize deployment control, data residency and inference customization over a fully managed API. On August 12th, Alibaba Cloud's Qwen team https://github.com/QwenLM/Qwen3.8?ref=runtimewire released Qwen3.8-2.4T-A95B /models/qwen/qwen3.8-2.4t-a95b , an open-weight mixture-of-experts model with 2.4 trillion total parameters and 95 billion active during inference. Developers can download its weights and configuration files rather than relying exclusively on a hosted model provider. The release extends Alibaba's push to put a Qwen-Max-class model into the hands of infrastructure teams willing to operate it. Alibaba https://www.sec.gov/Archives/edgar/data/1577552/000095017025090161/baba-20250331.htm?ref=runtimewire , founded in 1999 by Jack Ma and 17 co-founders, develops Qwen through its Alibaba Cloud business rather than through a separately financed startup. A very large model with local deployment options The Qwen3.8-2.4T-A95B model card https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B?ref=runtimewire describes a 92-layer architecture with 512 experts, 10 routed experts and one shared expert active in each mixture-of-experts block. Its native context length is 262,144 tokens, extensible to 1,010,000 tokens. Alibaba calls it the first Qwen-Max-class model released with open weights. The text-only model requires thinking mode, placing its reasoning inside <think tags before the final answer. Operators can adjust reasoning depth with the reasoning effort setting at low, medium or xhigh levels. A separate preserve thinking option carries reasoning context across prior messages and is enabled by default. The downloadable artifacts support Hugging Face Transformers and serving engines including vLLM, SGLang and TokenSpeed, according to the model card. Alibaba also offers managed access through Qwen Cloud https://www.qwencloud.com/?ref=runtimewire , where Qwen3.8-Max /models/qwen/qwen3.8-max adds features including vision input, non-thinking operation, built-in tools and a one-million-token default context. Local control comes with a substantial infrastructure requirement. Activating 95 billion parameters for each token still demands expensive accelerator capacity, memory and serving expertise. The open-weight release gives operators control over deployment and inference behavior; it does not make a model of this size inexpensive to run. Open weights versus a hosted reasoning model The clearest contrast is with OpenAI's proprietary GPT-5.5 Pro /models/openai/gpt-5.5-pro . OpenAI announced GPT-5.5 https://openai.com/index/introducing-gpt-5-5/?ref=runtimewire on April 23rd, 2026. The original launch text said GPT-5.5 and GPT-5.5 Pro would reach the API "very soon," while a dated April 24th update at the top of the same announcement says both models were then available in the API. OpenAI describes GPT-5.5 Pro as a higher-compute reasoning model for demanding coding, research and knowledge work. OpenAI's API documentation https://developers.openai.com/api/docs/models/gpt-5.5-pro?ref=runtimewire describes it as supporting a 1,050,000-token context window. The two distribution models create different tradeoffs. GPT-5.5 Pro customers buy access to OpenAI's hosted model and infrastructure. Qwen customers can download the weights, choose a serving engine, modify the inference path and keep prompts within their own environment. Managed Qwen Cloud access remains available for teams that prefer Alibaba to handle deployment. For engineering leaders, the practical comparison extends beyond benchmark scores. Hardware availability, serving throughput, tool support, data residency and the ability to inspect or change inference behavior can determine whether a model fits a production workload. Qwen's open weights broaden those choices while shifting more operational responsibility to the customer. Reasoning controls create their own security work Qwen3.8's mandatory thinking mode also puts reasoning-trace handling on the deployment checklist. Applications need policies governing whether traces are retained, exposed to users, written into logs or passed between agents. Those records can contain material absent from the final response. The separate Stolen Thoughts research project https://stolen-thoughts.com/?ref=runtimewire documented the security risk around stored reasoning traces. Its authors said they collected 6,708 publicly available agent trajectories from GitHub and Hugging Face and reconstructed 315,320 encrypted reasoning blocks. In genuine user sessions, they found 704 distinct privacy artifacts, including API keys, passwords, access tokens and personal email addresses. Sixty-four appeared only inside reasoning blocks, according to the researchers. That research covered trajectories produced by Claude, GPT and Gemini models and does not allege a Qwen vulnerability. It does show why teams running models with explicit or retained reasoning need to treat those traces as sensitive application data. Alibaba packages Qwen for both camps Qwen3.8 competes for coding, research and agent workloads against hosted systems and other open-weight releases. Alternatives include DeepSeek V4 https://deepseek.com/en/transparency/?ref=runtimewire , Google's hosted Gemini 3.1 Pro https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/?ref=runtimewire and MiniMax M2.5 https://www.minimax.io/news/minimax-m25?ref=runtimewire . Alibaba is pairing downloadable models with a commercial cloud layer. RuntimeWire previously reported that QwenCloud combines Alibaba's models, agent tools and compatible APIs /article/alibaba-qwencloud-eddie-wu-ai-native-cloud-strategy , while a later release gave Qwen3.8-Flash a one-million-token API option /article/alibaba-qwen3-8-flash-api-qwencloud-pricing . The 2.4-trillion-parameter release gives operators another frontier-scale model they can place on infrastructure they control. The bill for that control arrives in GPUs, memory, deployment engineering and security practices rather than solely as an API invoice.