Alibaba releases Qwen3.8-Flash-Next: activates only 6B parameters per token with 50% price cut
智东西 · wechat · 2026-08-27
Alibaba's Qwen team open-sourced Qwen3.8-Flash-Next, a preview based on the next-generation Qwen4 architecture. With 125B total parameters, the MoE model activates only 6B parameters per token, reducing training costs to 1/9th of its predecessor. Key innovations include Qwen Sparse Attention (QSA), Gated Residual (GR), and N-gram Embedding. It supports a native 262k context window, extendable to 1 million tokens. It outperforms DeepSeek-V4-Flash and Claude-Opus-4.6 on multiple benchmarks. API pricing is set at 1 RMB input / 3 RMB output per million tokens, 33% lower than competitors.
More from Models
- China's StepFun releases Step 5 Preview weights: 600B MoE, 27B active, 1M context — Dr_Singularity · 2026-10-15
- Google says teams across the company are using Gemini 4 Argon on its hardest engineering problems — CurieuxExplorer · 2026-10-15
- Perplexity's new embedding models: index with a 9B model, search with 0.6B — antoine_chaffin · 2026-10-15
- 797GB of n-gram counts in CPU RAM let 1B models match 57% more compute — PangWeiKoh · 2026-10-15
- Leak: Moonshot's Kimi flagship 'x4' done with pre-training, K3.1 in internal testing — lyraxana · 2026-10-15
- Unsloth ships GGUF quants for Nemotron 3.5: Bolt-74B runs in 24GB RAM — danielhanchen · 2026-10-15