Alibaba released Qwen3.8-Flash-Next on Aug. 26, an open-weight model that previews the architecture behind the coming Qwen4 family and activates 6 billion parameters per token.
What to Know:
Qwen3.8-Flash-Next pairs a 125 billion parameter backbone with a 51 billion parameter lookup table, but fires only 6 billion parameters on each token.
Alibaba says the model cost about one-ninth as much to train as its 397 billion parameter predecessor.
The model reads 262,144 tokens natively and stretches to 1 million, with weights posted publicly under a community license.
Qwen3.8-Flash-Next Specs and Cost
The Qwen team published the weights on Hugging Face and ModelScope, and it framed the drop as an early look at the design its next flagship family will inherit. The checkpoint carries a 125 billion parameter main model, a 51 billion parameter N-gram embedding table and a 4 billion parameter prediction module. Alibaba had announced the timing a day ahead.
Only 6 billion parameters switch on for any given token. The FP8 file still weighs about 173 gigabytes, so self-hosting takes a multi-GPU server rather than a workstation.
The model handles 262,144 tokens out of the box, and it reaches 1 million tokens through a scaling method called YaRN. Alibaba puts the training bill at roughly one-ninth of Qwen3.7-Plus, a 397 billion parameter model that activates 17 billion parameters per token.
Also Read: AI Agent Payments Breakthrough: x402 Protocol Launches Faster Transactions
Qwen4 Design Wins NVIDIA Support
Three of every four layers use Gated DeltaNet, which compresses past context into a fixed recurrent state instead of a cache that keeps growing. The fourth layer runs Qwen Sparse Attention, a retrieval step that scores blocks of text rather than single tokens. The pattern repeats 12 times across 48 layers.
NVIDIA validated the model on its GB300 NVL72 rack the day the weights appeared, reporting more than 16,000 tokens per second per GPU. Alibaba’s own kernel tests claim prefill speedups of up to 7.6 times at a 1 million token context.
Those numbers come from the vendors. Reviewers who dissected the release urged caution until outside teams reproduce the gains, and they flagged the qwen-community-1.0 license as a term commercial users should read carefully.
Why Alibaba’s Cheap Tokens Matter
Alibaba’s published benchmarks put Qwen3.8-Flash-Next ahead of Anthropic‘s Claude Opus 4.6 on software repair, scoring 62.5 on SWE-bench Pro against 53.4. The gap widens on office work, where the model reached 73.9 on CoWorkBench while DeepSeek-V4-Flash managed 45.1.
Frontier reasoning still favors rivals. Claude Opus 4.6 leads on Humanity’s Last Exam, 40.0 to 35.9, and DeepSeek-V4-Flash wins the NL2Repo test by six points.
The hosted version charges $0.16 per million input tokens and $0.47 per million output tokens, roughly a twelfth of what the Qwen3.8-Max flagship costs. That math keeps pressure on American labs selling similar capability at higher rates.
The release follows a pattern Alibaba set a year ago. It shipped Qwen3-Next as an architecture preview before the Qwen3.5 family, then carried that hybrid attention design through four model generations. It introduced Qwen3.8-Max as its flagship in early August, weeks after opening the weights of the smaller Qwen3.8-27B under a permissive license.
Read Next: Perplexity’s Proven Fix for the AI Search Gap: 500 Million Records