FULL STORY

Qwen3.8-Flash-Next Open-Sources a Preview of Qwen4 Architecture

Alibaba's Qwen team open-sourced the multimodal MoE model Qwen3.8-Flash-Next, a preview of the Qwen4 architecture. The rollout spanned the announcement, a detailed technical report, FP8 and community variants, and benchmarks that fueled wide discussion.

2026-08-25 ~ 2026-08-30 · 7 episodes · 57 posts

Episode 1 · Alibaba's Qwen Team Announces Qwen3.8-Flash-Next as a Preview of Qwen4 Architecture (2026-08-25, 13 posts)

On August 25, Alibaba's Qwen team announced via ModelScope that its next-generation Qwen4 architecture is ready, and launched a countdown page on Hugging Face for Qwen3.8-Flash-Next, confirmed for open release on August 26. Positioned as "a preview of the Qwen4 architecture," the release drew strong community attention, with notify-me signups growing from 45 (@NielsRogge, Aug 25) to 515 (@victormustar, Aug 26).

Confirmed

  • Official announcement that Qwen3.8-Flash-Next opens on August 26; the preview page shows August 26, 2026
  • Open-weight multimodal mixture-of-experts (MoE) design
  • 125B total parameters, including 51B N-gram parameters, with only 6B activated per token
  • Built on the next-generation architecture that will underpin the upcoming Qwen4 family
  • On Aug 26, @alexcovoeth relayed that the Qwen3.8-120B/51B/A6B MoE series would launch within 24 hours
  • @TheZachMueller relayed that Qwen3.8 Flash is optimized for a single DGX Spark and can run in NVFP4 precision

Unconfirmed

  • Whether @alexcovoeth's relay of a "Qwen3.8-120B/51B/A6B series within 24 hours" means additional models ship alongside remains officially unconfirmed
  • The exact role of the 51B N-gram lookup parameters; superelesha's analysis (in @teortaxesTex's thread) awaits post-release testing
  • DGX Spark deployment details come from third-party relay; precision and configuration await official confirmation

Why it matters

  • Per @calabiandyau, this effectively lets developers try out Qwen4's architectural improvements early, marking another major Qwen update
  • @rohanpaulai notes the large-total/small-active design targets high inference efficiency, potentially a new cost-effective open-model option if performance delivers
  • The 51B N-gram lookup/embedding parameters are a rare configuration among mainstream open models, drawing attention from community members like @zephyrz9
  • If the single-DGX-Spark rumor holds, local deployment barriers drop further

Episode 2 · Alibaba Open-Sources Qwen3.8-Flash-Next, an Early Preview of the Qwen4 Architecture (2026-08-26, 26 posts)

On August 26, Alibaba's Qwen team released and open-sourced the multimodal MoE model Qwen3.8-Flash-Next on Hugging Face and ModelScope, positioning it as an early preview of the Qwen4 architecture, accompanied by a technical report and detailed evaluations. The model has 125B main parameters plus roughly 51B N-gram embedding parameters, activating only 6B per token; officials and media claim training costs dropped to about 1/9 of the previous generation, with Day-0 support from vLLM, SGLang, and NVIDIA NeMo. Community reports say it beats DeepSeek V4 Flash with about half the parameters and surpasses Claude Opus 4.6, topping ClawEval and multilingual SWE-Bench—making it the most-watched open-source release of the day, with coverage from QbitAI and Zhidongxi spreading the news from August 27.

Confirmed

  • Official accounts (@AlibabaQwen, @Qwen) confirmed the multimodal MoE architecture: 125B total main parameters, 6B activated per token, with an extreme cost-efficiency pitch (media citing halved pricing and 1/9 training cost).
  • Architecture upgrades include GDN+QSA hybrid attention, GatedResidual multi-branch residuals, and N-gram Embedding capacity expansion (51B extra N-gram embedding parameters); community posts relayed details of 36 Gated DeltaNet layers plus 12 Qwen Sparse Attention layers, with 23.5GB memory savings claimed.
  • The model supports image-text-to-text input, safetensors format and API compatibility, and an FP8 open-weight release on Hugging Face.
  • Pricing and context: Qwen3.8-Flash is live on QwenCloud with native 262K-token context extensible to 1M tokens. Domestic API pricing per million tokens: ¥0.8 input, ¥2.7 output, ¥0.1 cache hit (flat rate per @AlchainHust); overseas: $0.15 input, $0.47 output, $0.016 cache hit per @MaziyarPanahi.
  • Day-0 ecosystem support: vLLM, SGLang, and NVIDIA NeMo (NeMo AutoModel and NeMo RL for domain fine-tuning).
  • @jiqizhixin confirmed availability on Hugging Face and ModelScope with a technical report and evaluations.
  • Benchmarks (community-relayed): Agents' Last Exam score 51.2, Pass@1 24.3%; ranked first on ClawEval and multilingual SWE-Bench, second on DeepSWE, third on SWE-Bench Pro; @bindureddy later claimed it beats GLM 5.3 Flash and DeepSeek Flash, calling it the best Flash model on the market.
  • @sammcj opened a discussion thread covering quantization, fine-tuning, chat templates, and inference server support.

Unconfirmed

  • @ariG23498 cited 180B parameters, inconsistent with the official 125B + 51B breakdown; treat the official figure as authoritative.
  • All comparative claims (beating Claude Opus, DeepSeek V4 Flash, GLM 5.3 Flash) and rankings come from community relay posts; layer-structure and memory-saving details remain unverified against the official report. Superlatives such as "lowest cache-hit price globally" are author opinions; @bindureddy's $0.16 input price differs slightly from @MaziyarPanahi's $0.15.

Why it matters

  • As the first preview of the Qwen4 architecture, its ultra-sparse MoE design (6B/125B activation) and new components (GDN+QSA, GatedResidual, N-gram Embedding) signal Alibaba's next-generation direction, with N-gram Embedding echoing DeepSeek's Engram idea—an open-source architecture relay.
  • Radically lower costs plus open weights and rapid ecosystem support could directly challenge premium closed models and competing Flash-tier models on price-performance, provided the benchmark claims hold up.

6 more related posts →

Episode 3 · Qwen3.8-Flash ships with local deployment and claims SWE-bench Pro win over Claude (2026-08-26, 5 posts)

Alibaba Qwen released Qwen3.8-Flash (later Qwen3.8-Flash-Next), an open-weight multimodal MoE model on a new "Next" architecture, widely seen as a preview of Qwen4. Unsloth enabled local running with just 75GB RAM and shipped 1-4bit GGUF quantizations, recommending llama.cpp. Community tests report 62.5 on SWE-bench Pro, beating Claude Opus 4.6 Max's 53.4 at roughly 1/9 the inference cost.Confirmed - 125B total parameters, multimodal MoE with only 6B active parameters per token - 51B N-gram Embedding attached to the main model (per @AGI Hunt) - Unsloth local deployment (75GB RAM) and 1-4bit GGUF quantizations via llama.cpp (also relayed by @MaziyarPanahi) - @Hesamation tested 62.5 on SWE-bench Pro vs Claude Opus 4.6 Max's 53.4Unconfirmed - Benchmark scores and the "1/9 cost" claim come from the vendor and community relays; no independent verification yet - @Hesamation says the model previews Qwen4 architecture; official Qwen4 details remain unknownWhy it matters - An open-weights model claims to beat top closed-source models on a real engineering benchmark at a fraction of the cost - Sparse MoE plus extreme quantization makes a 125B-class model locally deployable on consumer/workstation hardware

Episode 4 · Qwen3.8-Flash-Next FP8 Quantized Model Released on Hugging Face (2026-08-27, 2 posts)

Alibaba's Qwen team and unsloth released the Qwen3.8-Flash-Next-FP8 quantized model on Hugging Face, based on Qwen3.8-Flash-Next with FP8 quantization.

Episode 5 · Qwen 3.8-Next Released with Detailed Technical Report (2026-08-27, 2 posts)

Qwen 3.8-Next has been released with a detailed technical report covering architecture and pretraining experiments. Analysts predict that hybrid attention combining global and linear/sparse attention will become mainstream for large open-source models.

Episode 6 · Uncensored Qwen3.8-Flash-Next Weights and GGUF Released (2026-08-27, 2 posts)

OrcaRouter released uncensored Qwen3.8-Flash-Next weights supporting 262K context in MLX and GGUF formats for safety research and red-teaming, with a GGUF version now also available on Hugging Face.

Episode 7 · Alibaba Open-Sources Qwen3.8-Flash: 125B MoE with 6B Active Params, 20x Cheaper (2026-08-28, 7 posts)

On August 28, Alibaba's Qwen team released and open-sourced Qwen3.8-Flash, the third model in the Qwen3.8 lineup after Max and 27B. It uses a new-generation MoE architecture with 125B total parameters but only 6B active per token, cutting inference cost roughly 20x versus Qwen3.8 Max, and topped Hugging Face leaderboards upon release.

Confirmed

  • Specs: 125B total / 6B active parameters, native 262k context extendable to 1M tokens, multimodal; available in OpenCode Go out of the box; open-source and free.
  • Cost: $0.16 per million input tokens, training cost down nearly 90% from the previous generation; @VraserX claims training cost is about one-ninth of Qwen3.7-Plus, arguing the collapse in the cost of creating intelligence is broadly underestimated.
  • Performance: @大模型之路 reports it beats Claude Opus 4.6 on coding and office tasks; 量子位 notes a single RTX 4090 can run its 1M-token context.
  • Architecture: per a repost by @rsasaki0109, the model integrates N-gram Embedding and sparse attention among recent techniques (via repost, original source unverified).

Why it matters

  • Tiny active parameters plus 1M context slash deployment barriers, putting long-context multimodal models within reach of consumer hardware.
  • Channels like Two Minute Papers highlight its "free model rivaling billion-dollar ones" narrative, underscoring open-source pressure on closed-model pricing.