Descriptions:
Fahd Mirza takes Qwen3.8-Flash-Next for a hands-on spin the day it drops — an open-weight preview of the upcoming Qwen 4 architecture from Alibaba’s Qwen team. The model packs 125 billion total parameters but activates only 6 billion per token through a mixture-of-experts design, and Mirza runs the quantized GGUF version via llama.cpp on a single Nvidia H100 (80 GB VRAM), reporting just over 61 GB consumed at load.
The architectural highlights Mirza explains include Gated Delta layers that maintain a compressed running memory of the conversation context, Qwen Sparse Attention that searches history in chunks rather than scanning token-by-token, a gated residual system to preserve information depth, and an Engram embedding layer — essentially a large lookup dictionary of common word patterns held in RAM to boost knowledge without inference overhead. Together these let the model run at a fraction of the cost of a dense 125B model.
For evaluation, Mirza runs a demanding single-shot HTML generation task requiring nested tabbed UI, multilingual native-script country names, and broad factual coverage across continents — the model produces a working output. He also benchmarks it against Qwen-3.27B, Qwen-3.7+, DeepSeek V4 Flash, and Claude Opus 4.6, finding Flash-Next competitive or ahead on most language and coding benchmarks while losing on HLE multidisciplinary reasoning (Claude Opus 4.6 leads) and repo-level code generation (DeepSeek V4 Flash leads). A follow-up API test via Qwen’s hosted endpoint covers multimodal and agentic vision tasks.
📺 Source: Fahd Mirza · Published August 26, 2026
🏷️ Format: Review
