Qwen3.8 Flash (CanopyWave)

A third-party guide cross-checks Qwen3.8-Flash against official model cards, describing Flash-Next as a 125B MoE with 6B active per token across 512 experts, 51B n-gram embeddings, and a 4B MTP head, with 262,144 native tokens extendable to 1M via YaRN while the managed Qwen3.8-Flash defaults to 1M context. It emphasizes that Flash-Next ships under the Qwen Community License 1.0 rather than Apache 2.0, unlike the 27B dense anchor. For self-hosting, the guide notes the full checkpoint is 335 GB (FP8 173 GB), with the 51B n-gram table designed to live in host memory with async prefetch, and suggests the managed qwen3.8-flash API at $0.16 per million input tokens and $0.47 per million output tokens as the cheaper way to evaluate before budgeting.

An unofficial InferenceX profile for Qwen3.8-Flash-Next lists 48 layers with 36 Gated DeltaNet and 12 Qwen Sparse Attention layers, Top-10/513 MoE routing, a 2,560-dimensional token embedding with a 248,320 vocabulary, and 262K context. It notes that the widely quoted 125B figure covers only the main model, since a 51B n-gram embedding table brings the total to 176B, with a separate 4B multi-token prediction head outside both numbers. The page describes weights shipping under the qwen-community-1.0 license in BF16 and in an FP8 checkpoint using fine-grained quantization with a block size of 128, and states that thinking mode is enabled by default, emitting think blocks that can be tuned via enable thinking, preserve thinking, and reasoning effort settings. It flags that the deployment is Alibaba's open-weight preview of the architecture underlying Qwen4.

An arXiv technical report (2608.30320, submitted 31 Aug 2026) by Zihan Qiu and 35 co-authors describes the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and 51B parameters of n-gram embedding tables held off the accelerator. On fourteen pre-training benchmarks it leads the 397B-A17B predecessor on eight and trails it by at most 2.6 points. The paper says the model achieves that result at one third the activated parameters, one third the training tokens, and roughly one ninth the training FLOPs of its predecessor. Token mixing uses a layer-wise hybrid of Gated DeltaNet and global attention with one full-attention layer in every four, replaced at continued-pretraining time by Qwen Sparse Attention using micro-block context scoring with a compressed lightweight indexer.

QwenCloud's documentation for model ID qwen3.8-flash describes a next-generation architecture as an early preview of Qwen4, with 125B total parameters and only 6B activated per token, trained at less than one ninth the cost of Qwen3.7-Plus. It natively supports 1M context and multimodal understanding of text, images, and video. The docs detail systematic upgrades across attention, residual, embedding, and optimization: a GDN plus Qwen Sparse Attention hybrid claiming a 7.6x prefill speedup at 1M tokens, a Gated Residual that expands the information stream into four parallel branches, 51B n-gram embedding parameters at near-zero extra compute, and the Muon optimizer. Reported results include SWE-bench Pro 62.5 versus Claude Opus 4.6 at 53.4, DeepSWE 58.7, CoWorkBench 73.9, JobBench 55.7, and Toolathlon 73.5.

Alibaba released Qwen3.8-Flash, an open-weight multimodal Mixture-of-Experts model built on a 125B-parameter main model plus 51B n-gram embeddings, with 6B parameters activated per token. It is positioned as an early architectural preview of the upcoming Qwen4 series and targets high-volume applications, tool-driven workflows, and coding or co-working assistants. The model natively supports 262K tokens of context, extendable to 1 million tokens, and competes with DeepSeek-V4-Flash and Claude-Opus-4.6 on benchmarks including SWE-bench Pro, CoWorkBench, Toolathlon Verified, MathVision, AndroidWorld, and ERQA. Weights are available on Hugging Face and ModelScope, with API access on Model Studio and Qwen Cloud priced at 1 RMB input and 3 RMB output per million tokens.