Qwen3.8-Flash and Qwen3.8-Flash-Next launched August 26, 2026 as open-weight models and early previews of the forthcoming Qwen4 architecture. Both are available via API on QwenCloud; Qwen3.8-Flash became available on OpenRouter on August 27. Alibaba also released an FP8 variant of Qwen3.8-Flash-Next alongside the main weights.
The models carry 125B total parameters plus a 51B N-gram embedding table, with only 6B parameters activating per token. The architecture uses GDN and QSA as a hybrid attention mechanism, gated residual connections providing four information-passing lanes between layers instead of one, and the Muon optimizer for training. The N-gram table is designed to sit in host memory with async prefetch rather than GPU memory, and only a quarter of the model's layers maintain a growing KV cache. The model supports a 262K native context window extensible to 1M tokens via YaRN.
Qwen states the model was trained at 1/9 the cost of Qwen3.7-Plus while outperforming it, with particular gains in coding and office tasks. Cited benchmarks include 62.5 on SWE-bench Pro, 84.5 on AndroidWorld, 58.7 on DeepSWE 1.1, and 95.7 on MathVision. UnslothAI stated the model outperforms Claude Opus 4.6 (Max). The QwenCloud API is priced at $0.16/1M input tokens and $0.47/1M output tokens.
SGLang, vLLM, UnslothAI, and TokenSpeed all provided day-0 infrastructure support. UnslothAI's GGUF quantizations enable local execution on 75GB of RAM. The release reached the top of Hacker News with 272 points.