80B total / approximately 3B active per token (sparse MoE; 79.674B main-model learned parameters, 81.325B packaged including MTP)
262,144 tokens native (256K); up to 1,010,000 with explicit YaRN scaling
Apache 2.0
Original text-only Qwen3-Next release with two distinct post-trained checkpoints: Instruct is non-thinking only, while Thinking always uses reasoning and its chat template pre-fills <think>; these are not an enable_thinking toggle on one checkpoint. The 48-layer hybrid alternates three Gated DeltaNet layers with one gated full-attention layer (36 linear / 12 full), each followed by sparse MoE with 512 routed experts, 10 selected per token and one shared expert. Hidden width is 2,048; full attention uses 16 query / 2 KV heads of dimension 256 with partial RoPE, while DeltaNet uses 16 QK / 32 value heads of dimension 128. Qwen reports pretraining on a uniformly sampled 15T-token subset of its 36T-token Qwen3 corpus, followed by post-training; GSPO is explicitly credited for stabilizing Thinking RL, but the complete checkpoint-specific SFT/RL recipe and trained context schedule are not disclosed in the checked launch sources. Zero-centered RMSNorm, weight decay on normalization weights and normalized router initialization improve training stability. Multi-token prediction is trained to improve model quality and supplies an auxiliary speculative-decoding module; acceleration requires supported inference implementations and is not generally available in Transformers. Qwen claims base-model pretraining used 9.3% of Qwen3-32B compute cost (GPU hours) and that prefill/decode throughput exceeds Qwen3-32B by 10x beyond 32K context; the published comparison does not disclose hardware, precision or batch size and is not an end-to-end cost or universal latency guarantee. Publisher benchmarks report Instruct/Thinking GPQA scores of 72.9/77.2, AIME25 of 69.5/87.8 and LiveCodeBench v6 (February-May 2025) of 56.6/68.7; exact score-specific sampling and tool protocols are not fully specified, so these are not verified no-tool results. Instruct approaches the much larger Qwen3-235B-A22B-Instruct-2507 on selected tasks and improves long-context RULER, but does not match it uniformly; Thinking can require substantially longer generation. Native context is a shared input/output budget; the advertised approximately 1M extension requires YaRN factor 4 and an inference length override, with possible short-context degradation. Original downloads are BF16, with the full approximately 80B weights needed despite sparse activation; official FP8 packages appeared later and are not same-day launch variants. The evaluated original Base is not verified as a public launch checkpoint, and later Coder-Next and other derivatives are excluded. September 11 is provisional: official GitHub release news and contemporary local-use evidence support availability by that date, but the official download-linked blog displays September 10; exact first public access and the source calendar/timezone discrepancy remain unresolved.