GitHub radar
Qwen3.8-Flash-Next: Preview of Qwen4 Architecture
Alibaba's Qwen team released an experimental model that previews the architecture behind Qwen4 — 125B total parameters with only 6B active during inference, plus vision and coding capabilities.
Qwen3.8-Flash-Next is an experimental model from Alibaba's Qwen team that previews the architecture planned for Qwen4. It uses a radically different design: 125B total parameters but only 6B are activated at inference time, making it fast despite its size. The model combines hybrid attention (Gated DeltaNet + sparse attention), n-gram embeddings instead of dense token embeddings, and multi-token prediction. It's multimodal — handles text and images natively — with a 262K native context window (extendable to 1M). Benchmark highlights: SWE-bench Pro 62.5%, AndroidWorld 84.5%, GPQA Diamond 91.7%. The model can run via Ollama with `ollama run qwen3.8-flash-next`, though serious hardware is required.
Why a vibe-coder should care
This is the clearest signal yet of where Qwen — and by extension the open-source model ecosystem — is heading. The sparse activation design means comparable performance to Qwen3.8-27B at a fraction of the inference compute. If the Qwen4 release follows this architecture, quantized versions could become practical on high-end consumer hardware.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Read the model card at https://huggingface.co/Qwen/Qwen3.8-Flash-Next and explain in plain terms what's architecturally new compared to regular Qwen3.8 and why it matters.
This is a server-grade model — you can't run it at home: even the 1-bit compressed version takes 72 GB or more. Think of it as news about the future of Qwen, not a tool for right now.
Open on Hugging Face▌ More finds