AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Large Language Models

Qwen3.8-Flash-Next: Alibaba Releases 125B MoE With 6B Active Parameters as Qwen4 Architecture Preview

Alibaba open-sourced Qwen3.8-Flash-Next, a 125B multimodal MoE with only 6B active parameters, claiming near-frontier performance and training cost at one-ninth of the previous generation. The release previews the Qwen4 architecture.

AI News Feed · 2026-08-27 03:29

Alibaba reports that Qwen3.8-Flash-Next outperforms Claude Opus 4.6 on several agentic and multimodal benchmarks while cutting training cost to roughly one-ninth that of Qwen3.7-Plus. The inference price is 1 yuan per million input tokens and 3 yuan per million output tokens, which the company says is two-thirds of DeepSeek-V4-Flash's off-peak price and one-third of its peak price. Leiphone said the model's overall efficiency sets a global benchmark.

The model combines a 125B backbone with a 51B n-gram embedding table and a 4B multi-token prediction module, totaling 180B on disk, with 6B parameters active per token. Four architectural changes drive the release: a hybrid of Gated DeltaNet and Qwen Sparse Attention (GDN in three of every four layers, QSA in the fourth), Gated Residual widening the residual stream into four parallel branches, N-gram Embedding for capacity via deterministic lookups, and the Muon optimizer combined with AdamW. Native context length is 262,144 tokens, extensible to 1,000,000 with YaRN.

On benchmarks, the Qwen team says Qwen3.8-Flash-Next scores 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 55.7 on JobBench, 73.5 on Toolathlon Verified, 84.5 on AndroidWorld, and 95.7 on MathVision with code interpreter. However, independent coverage by MarkTechPost notes that the model does not lead everywhere: Claude Opus 4.6 Max scores 40.0 on HLE versus Qwen's 35.9, and DeepSeek-V4-Flash-0731 leads NL2Repo-Bench at 54.2 versus 48.1. Frontier reasoning remains a gap.

Deployment requires a multi-GPU node rather than a workstation. The FP8 checkpoint is 172.78 GiB and the BF16 checkpoint is 335.28 GiB. The model runs under vLLM, SGLang, TokenSpeed, transformers serve, and llama.cpp; fine-tuning is supported via Unsloth, Swift, and LLaMA-Factory. It is licensed under qwen-community-1.0, not Apache-2.0, so commercial users should verify terms. The model already powers the "Standard" mode on QwenWork and works with Qwen Code.

The weights are open-sourced on Hugging Face and ModelScope. The Qwen3.8 family now includes Qwen3.8-Max, Qwen3.8-27B, and Qwen3.8-Flash. According to Qwen team figures cited by Leiphone, global downloads of Qwen models have surpassed 3 billion, with more than 300,000 derived models.