CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
← Back to feed

Qwen3.8-Flash-Next Is a Qwen4 Architecture Preview — 125B MoE, Sparse Attention, Ships August 26

Alibaba’s Qwen team posted Qwen3.8-Flash-Next to ModelScope on August 25 with a countdown to August 26, 11:00 EDT. The model description is explicit: this is not a standard Qwen 3.8 series addition. It is a structural preview of the Qwen4 architecture, released early to let the community adapt tooling before the full family ships.

What Is Qwen3.8-Flash-Next

  • 125B total parameters, 6B active — standard MoE ratio, optimized for inference throughput
  • Multimodal — handles text and image inputs
  • FP8 variant included at launch (Qwen3.8-Flash-Next-FP8)
  • Built on the Qwen4 architecture

The “Flash-Next” naming signals the intended use case: fast inference at efficiency, positioned below the Qwen3.8 Max flagship (2.4T parameters, released August 4) and above the Qwen3.8-27B dense model (released August 14 with 81.8k downloads in 11 days).

Qwen4 Architecture: Sparse Attention

A screenshot circulated on X shortly after the ModelScope page went live captured details that were quickly removed from the official description. The leaked text referenced Qwen Sparse Attention as the new attention mechanism, along with 51B n-grams as a component of the architecture.

Sparse attention reduces compute cost by attending to a selected subset of tokens rather than the full context at each layer. If Qwen Sparse Attention follows the pattern of other sparse attention designs (like MLA in DeepSeek V3/V4), it likely trades some precision for significant memory and latency gains at long context.

Alibaba’s stated intent is to give the community time to prepare for Qwen4 — inference engines, quantization tools, and serving infrastructure all require adaptation when an attention mechanism changes. Releasing a smaller model with the new architecture ahead of the flagship is the same pattern DeepSeek used with V3-Mini before V4.

Context: Qwen3.8 Series

The Qwen3.8 family has shipped in rapid succession:

  • Qwen3.8-Max (August 4): 2.4T MoE, 86.6% Terminal-Bench, PaperBench SOTA at 93.0
  • Qwen3.8-27B (August 14): 27B dense, Apache 2.0, SWE-bench Pro beats Claude Opus 4.6 Max, 81.8k downloads in 11 days
  • Qwen3.8-Flash-Next (August 26): 125B MoE, 6B active, Qwen4 architecture preview

The Flash-Next release suggests Alibaba is treating the Qwen3.8 namespace as a testbed for architectural experiments before committing them to the flagship Qwen4 series. With 41 registered waiters on ModelScope as of the announcement, broader community uptake will follow once the weights are downloadable.

No benchmark numbers have been published. Evaluation results will emerge in the hours after release.