Qwen just open-weighted a 2.4-trillion-parameter model. Nobody else has done this.
The wait is over. Qwen3.8-2.4T-A95B is on Hugging Face with weights, which makes it the first Qwen-Max-class model to ever go open. 2.4 trillion total parameters, 95 billion activated, 92 layers mixing Gated DeltaNet with gated attention, 512 experts per layer with 10 routed plus one shared. Native context is 262,144 tokens, extensible to 1,010,000. A dense 27B companion went out at the same time for people who do not own a datacenter.
The scores are not a courtesy release. SWE-bench Pro 67.7%, GPQA Diamond 92.6%, OneMillion-Bench expert score 52.5. That is frontier territory, published under an open license, downloadable tonight. FP8 weights are up, unsloth already has GGUF quants, and it runs on vLLM, SGLang and TokenSpeed out of the box.
Two design choices are worth staring at. Thinking mode cannot be turned off β every response begins with reasoning, and there is a preserve_thinking flag that carries reasoning context across messages instead of throwing it away each turn. For agent loops that is a real behavioral change: the model keeps its own scratchpad between steps. And reasoning_effort is exposed as xhigh, medium, low, so you pay for depth only when the task needs it.
Here is the part that matters beyond the leaderboard. Every Western lab that talked about openness this year shipped something in the 20B-to-30B range and called it a gift to the community. Alibaba just put its actual flagship on the table. Whatever you believe about the motives, the practical effect is that the strongest model you can run on your own hardware is now Chinese, and it is not close.
Weights at https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
← Back to all articles
The scores are not a courtesy release. SWE-bench Pro 67.7%, GPQA Diamond 92.6%, OneMillion-Bench expert score 52.5. That is frontier territory, published under an open license, downloadable tonight. FP8 weights are up, unsloth already has GGUF quants, and it runs on vLLM, SGLang and TokenSpeed out of the box.
Two design choices are worth staring at. Thinking mode cannot be turned off β every response begins with reasoning, and there is a preserve_thinking flag that carries reasoning context across messages instead of throwing it away each turn. For agent loops that is a real behavioral change: the model keeps its own scratchpad between steps. And reasoning_effort is exposed as xhigh, medium, low, so you pay for depth only when the task needs it.
Here is the part that matters beyond the leaderboard. Every Western lab that talked about openness this year shipped something in the 20B-to-30B range and called it a gift to the community. Alibaba just put its actual flagship on the table. Whatever you believe about the motives, the practical effect is that the strongest model you can run on your own hardware is now Chinese, and it is not close.
Weights at https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
Comments