Alibaba's Qwen3.8-Flash activates only 6B of 125B params, cuts costs by 90%
大模型之路 · wechat · 2026-08-28
Alibaba open-sourced Qwen3.8-Flash, featuring a next-gen MoE architecture with 125B total parameters but activating only 6B. It outperforms Claude Opus 4.6 on core tasks like coding and office work, reducing training costs by nearly 90% and offering an input price of $0.16 per million tokens.
Key Features
- Efficient Architecture: Introduces GatedResidual design to preserve cross-layer signals with low activation, preventing information collapse.
- Long Context: Natively supports 262K context, expandable to 1M via YaRN.
- Ecosystem: Weights open-sourced on HF and ModelScope; vLLM and SGLang support Day-0 inference.
Industry Impact
Marks a shift in the LLM race from "more parameters" to "higher efficiency." Flagship-level capabilities are now potentially viable on consumer hardware for local deployment, significantly lowering the barrier for small teams.
More from Models
- Unredacted Anthropic filing exposes internal data-quality debate months before Opus 3 cutoff — zacharynado · 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