📊 BENCHMARKS & EVALS
- Alibaba releases Qwen3.8-Flash-Next with native 262K-token context
Alibaba released Qwen3.8-Flash-Next, an open-weight multimodal model with a native 262K-token context.- For developers and AI teams building long-context, multimodal, agentic, and self-hosted applications.
- The Qwen4 architecture preview has 176B total parameters, 6B active per token, and a 262,144-token native context window.
- The open weights support Transformers, vLLM, SGLang, and llama.cpp; YaRN can extend context to approximately 1M tokens.
- Its hybrid Gated DeltaNet and sparse-attention design reduces long-context memory and attention costs; a 4-bit quantization ran at 11 tokens per second at 262K context on a 128GB M5 Max.