Alibaba releases Qwen3.8-Flash-Next, an open-weight Qwen4 preview

Alibaba's Qwen team announced the open-weight release of Qwen3.8-Flash-Next, a multimodal mixture-of-experts model that previews the Qwen4 architecture. Weights are on Hugging Face and ModelScope. A production version, Qwen3.8-Flash, runs on QwenCloud with a 1M-token context by default and built-in tools. It costs $0.15 per million input tokens and $0.47 per million output tokens.

The model has 176B total parameters: a 125B main model plus 51B N-gram embedding parameters. It activates 6B parameters per token. Native context is 262,144 tokens, extensible to 1M with YaRN. Alibaba says training cost about 1/9 as much as for Qwen3.7-Plus, yet the new model beats Qwen3.7-Plus at coding and office tasks.

Three of every four layers use Gated DeltaNet, which compresses history into a fixed-size recurrent state, so its KV cache does not grow as sequences get longer. The fourth uses Qwen Sparse Attention (QSA), which groups the sequence into micro-blocks and scores importance per block rather than per token. That cuts the indexing overhead that grows with context in earlier sparse-attention designs. Gated Residual also widens the residual stream into four gated branches, the N-gram embedding table can be offloaded to host memory, and training used the Muon optimizer.

Alibaba's own benchmarks put the model at 62.5 on SWE-bench Pro, ahead of Qwen3.7-Plus (55.8), DeepSeek-V4-Flash (56.0) and Claude Opus 4.6 (53.4). It also scores 81.0 on SWE-bench Multilingual, 58.7 on DeepSWE 1.1, 91.7 on GPQA Diamond, 35.9 on HLE, 91.9 on LiveCodeBench v6, 73.5 on Toolathlon Verified and 73.9 on Alibaba's in-house CoWorkBench. On vision tasks it scores 84.5 on AndroidWorld and 52.3 (partial score) on OSWorld 2.0.

NVIDIA's developer blog reports Alibaba's efficiency figures. The QSA attention kernel is up to 7.6x faster than full attention during prefill and 4.9x faster during decoding. In a serving test at 1M-token context with a 90% prefix-cache hit rate, the model reached 8.6x the prefill throughput of Qwen3.7-Plus.

NVIDIA says it provides best-effort Day-0 functional support through SGLang, vLLM and TensorRT LLM, plus post-training recipes through NeMo AutoModel and NeMo RL. It reports over 16K tokens/second per GPU and over 200 tokens/second per user on GB300 NVL72, a 72-GPU Blackwell Ultra rack with a 130 TB/s NVLink domain. The model also runs on DGX Station, DGX Spark clusters and four RTX PRO 6000 workstation GPUs.

All accuracy and speed figures are self-reported by Alibaba and NVIDIA. No independent evaluation has been published.