FULL STORY

Alibaba Open-Sources Qwen3.8-Flash-Next, Previewing Qwen4 Architecture

After announcing Qwen4's readiness on August 25, Alibaba open-sourced the 125B sparse MoE model Qwen3.8-Flash-Next the next day, offering a first preview of the Qwen4 architecture.

2026-08-25 ~ 2026-08-26 · 2 episodes · 24 posts

Episode 1 · Alibaba Announces Qwen3.8-Flash-Next, a Qwen4 Architecture Preview Set for August 26 Release (2026-08-25, 13 posts)

On August 25, Alibaba's Qwen team announced via ModelScope that its next-generation architecture, Qwen4, is ready, kicking off the countdown to the release of the open-source model Qwen3.8-Flash-Next, which officially launched the following day (August 26). The community widely sees the move as a prelude to the formal debut of the Qwen4 series.

Confirmed

  • Qwen3.8-Flash-Next was officially announced as open-source, releasing the day after the announcement
  • The model features an open-weight multimodal mixture-of-experts (MoE) design
  • It has a total of 125 billion parameters, including 51 billion N-gram parameters, with only 6 billion activated per token
  • The model is built on the next-generation architecture that will underpin the upcoming Qwen4 series

Why it matters

  • According to @calabiandyau, this move effectively lets developers get early hands-on experience with the architectural improvements of the Qwen4 family, marking another major update to the Qwen lineup
  • @rohanpaulai noted that the design of a large total parameter count paired with low active parameters aims for high-performance inference efficiency; if it delivers, it could become a new cost-effective option among open-source models
  • N-gram parameters are a configuration rarely seen in mainstream open-source models; community members including @zephyrz9 have expressed interest in this architectural detail, though its exact role remains to be verified after release

Episode 2 · Alibaba Open-Sources Qwen3.8-Flash-Next: 125B Ultra-Sparse MoE Previewing Qwen4 Architecture (2026-08-26, 11 posts)

On August 26, Alibaba's Qwen team released and open-sourced the multimodal MoE model Qwen3.8-Flash-Next on Hugging Face, positioned as an early preview of the Qwen4 architecture. The community quickly launched discussions around quantization, fine-tuning, inference deployment, and benchmarks, making it one of the most-watched open-source model events of the day.

Confirmed

  • The official accounts (@AlibabaQwen, @千问大模型) confirmed the model uses a multimodal MoE architecture with 125B total parameters and only 6B active per token, targeting extreme cost-efficiency — @千问大模型 cited a 9x cost reduction.
  • Architecture upgrades include GDN+QSA hybrid attention, GatedResidual multi-branch residuals, and N-gram Embedding for expanded capacity (an additional 51B n-gram embedding parameters).
  • The model supports image-text-to-text multimodal input, is compatible with the safetensors format and API, and an FP8 version is open-sourced on Hugging Face.
  • Several community members (@kimmonismus, @SimplyAnnisa) relayed benchmark results showing strong performance — @SimplyAnnisa reported wins in 8 of 9 comparison benchmarks, and @kimmonismus claimed it beat Claude Opus.
  • @sammcj opened a centralized discussion thread covering quantized variants, fine-tuning and de-tuning, chat templates, inference server configs, and model comparisons.

Unconfirmed

  • @ariG23498 stated the parameter count is 180B, inconsistent with the official 125B total + 51B n-gram embedding figure — possibly an alternative counting that sums both; the official release remains the source of truth.
  • Claims such as "6B active parameters beating Claude Opus" come from community relays; exact test conditions and official leaderboard details await the full technical report.

Why it matters

  • As the first preview of the Qwen4 architecture, its ultra-sparse MoE design (6B/125B activation ratio) and new components like GDN+QSA signal the technical direction of Alibaba's next-generation models.
  • The drastically reduced active parameters cut inference costs significantly; if the benchmark results hold up, it poses a direct cost-performance challenge to high-end closed-source models like Claude Opus.