PublishersSemiAnalysis

SemiAnalysis

1 document read from this publisher.

Qwen3.8-Flash-Next — Architecture, Evals & Inference Performance

technical report · third party analysis · 2026-08-26 · 17 techniques

Mixture of Experts Multi-Token Prediction Hybrid Attention Gated DeltaNet Gated Residual N-gram embedding lookup Qwen Sparse Attention Category-specific assignment of Muon and AdamW NEXTN speculative decoding 10 Routed + 1 Shared Experts Activated per Token Per-Branch Scalar Write Gate Training without batch-size warmup Elementwise Data-Dependent Residual Read Gate Fine-grained FP8 quantization Fixed MTP draft-token acceptance length Fixed-budget sparse-attention selection MQA indexer