PrototypeAI
Alibaba open-sources Qwen3.8-Flash at about one-ninth training cost
- Alibaba calls the 125B-parameter model a technical preview of its next-generation Qwen4 architecture.
- Only about 6B parameters activate per token, cutting compute while supporting 262K-token native context.
- FP8 weights are available on Hugging Face and ModelScope; API pricing starts at 1 yuan per million input tokens.
6 billion parametersParameters activated per token
0
0
1 source · Pandaily (en)