Is Qwen (Alibaba) Down?

QWEN-IMAGE-2.1-TURBO -- AN 8-STEP CHECKPOINT, SAME RESEARCH-ONLY LICENCE (Hugging Face repo created 2026-10-09, vendor-primary model card): 'Qwen-Image-2.1-Turbo is an accelerated checkpoint of Qwen-Image-2.1 for text-to-image generation and image editing with 8 denoising steps. It uses the same 7B visual generation architecture and loads directly with QwenImage21Pipeline in Diffusers.' It ships its own sampling schedule, uses CFG=1 by default and reuses text and reference-image context across steps with prefix KV caching. **Licence front matter: license_name qwen-research** -- the same non-commercial research licence as Qwen-Image-2.1, so the speed-up does not change what you may build on it. Requires a Diffusers build from source (PR #14950). No speed or quality numbers versus the full-step model are stated in the card's opening sections.

QWEN-IMAGE-2.1 IS OPEN-WEIGHT -- BUT UNDER A RESEARCH-ONLY LICENCE, WHICH IS THE FIRST THING TO CHECK BEFORE YOU BUILD ON IT (weights 2026-09-14, blog and licence 2026-09-20; vendor-primary via Hugging Face and Qwen's blog): Alibaba published **Qwen-Image-2.1**, 'a unified text-to-image generation and image editing model' whose visual generation component is **7B parameters across 32 Single-Stream DiT layers**, with a mixed-granularity attention scheme (token-level causal mask for text, chunk-level for images) and **prefix KV-cache reuse** so reference images and editing instructions are computed once and cached. **What is new in capability terms:** (1) **native transparency** -- it generates regular or RGBA images from a prompt, edits transparent layers, and extracts a subject from a photograph as an RGBA layer, folding the December 2025 Qwen-Image-Layered model into the main line; (2) **up to 10 reference images** in one edit (group portraits from six headshots, a full outfit from five product shots, a room from ten furnishings); (3) **local editing by circles, painted annotations, or a separate mask** passed alongside the untouched original; (4) identity and product-fidelity preservation, panoramas, infographics, storyboards, and improved typography and portrait lighting. Output up to 2048x2048, with preset ratios to 2752x1536. Two companion **prompt-rewriting models (Qwen-Image-2.1-PE-T2I and PE-I2I, 2026-09-20)**, fine-tuned from Qwen3.5-VL 9B, turn short requests in any language into detailed English prompts with a recommended aspect ratio. **BENCHMARKS: chart only.** The blog shows a 'Qwen-Image-Bench' comparison against open and closed models as an image with no table, so no scores are printed here. **THE LICENCE, VERBATIM FROM THE REPO:** the model ships under the **Qwen RESEARCH LICENSE AGREEMENT (release date September 20, 2026)**, which grants rights 'FOR NON-COMMERCIAL PURPOSES ONLY', defines Non-Commercial as 'for research or evaluation purposes only', and says 'You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us' (model-business@notice.qwencloud.com). **That is a departure from the Apache-2.0 image releases this page praised in 2025 and a different posture from Qwen3.8-27B (Apache 2.0)** -- so 'open-source', the word Qwen uses in its own headline, means open for research here, not open for products. **VERSIONING TRAP:** this page recorded a **Qwen-Image-3.0 (2026-07-21, Qwen Chat only, no weights)**; the open release carries the lower number 2.1. Treat 2.1 as the open-weight line and 3.0 as the hosted flagship rather than reading 2.1 as a downgrade or a successor. Early traction as of 2026-09-21: about 6,500 downloads and 1,400 likes on the Hugging Face repo, one day after the blog post.

STALENESS CATCH-UP -- THREE QWEN3.8 OPEN-WEIGHT RELEASES IN AUGUST, INCLUDING THE FIRST OPEN MAX-CLASS MODEL, AND THE QWEN4 ARCHITECTURE PREVIEW (2026-08-05, 2026-08-08 and 2026-08-24/26; vendor-primary via Qwen's blog and Qwen's own Hugging Face model cards; this page was last reviewed 2026-07-22 and had only the WAIC preview): **(1) Qwen3.8-27B (weights on HF 2026-08-05, Apache 2.0)** -- the dense successor to the Qwen3.6-27B this page already recommends for single-GPU self-hosting; FP8 build followed 8/13. **(2) Qwen3.8-2.4T-A95B (weights on HF 2026-08-08, custom 'qwen3.8-max' licence, NOT Apache)** -- 'For the first time, Qwen3.8 brings a **Qwen-Max-class model to open release**': a 2.4T-parameter MoE with 95B active, the open counterpart of the hosted **Qwen3.8-Max** (which adds vision input, non-thinking mode, 1M context by default and built-in tools on QwenCloud). Reasoning depth via `reasoning_effort`; `preserve_thinking` keeps prior reasoning context. **The licence is the catch:** it is a bespoke Qwen3.8-Max licence rather than Apache 2.0, so read it before commercial deployment -- this is the same pattern as Kimi's K3 licence. **(3) Qwen3.8-Flash-Next (weights 2026-08-24, blog 2026-08-26, 'Qwen Community 1.0' licence)** -- 'a multimodal MoE model that also serves as an **early preview of the architecture used in Qwen4**', playing the role Qwen3-Next played for Qwen3.5. Four architectural changes: **Gated DeltaNet + Qwen Sparse Attention (QSA)** hybrid, **Gated Residual** (4-branch residual stream with dynamic gates), **N-gram Embedding** (a 51B lookup table that can be offloaded to host memory), and the **Muon optimizer**. Shape: **125B main model + 51B n-gram embeddings, 6B active per token**; 262,144 native context, 1M with YaRN. Qwen says training cost about **1/9 of Qwen3.7-Plus** while beating it on coding and office tasks. **Vendor benchmarks vs Qwen3.8-27B / Qwen3.7-Plus / DeepSeek-V4-Flash-0731:** DeepSWE 1.1 **58.7** / 42.2 / 16.5 / 54.4; SWE-bench Pro **62.5** / 61.7 / 55.8 / 56.0; SWE-bench Multilingual **81.0**; NL2Repo 48.1 (DeepSeek 54.2 leads); CoWorkBench **73.9**; JobBench **55.7**; Toolathlon Verified **73.5**; IFBench 81.3. All first-party. **Hosted pricing:** the production version, **Qwen3.8-Flash** on QwenCloud (1M context default, built-in tools), is **$0.15 per 1M input / $0.47 per 1M output** -- undercutting DeepSeek's new V4.1-Flash on output ($0.60 off-peak) and matching it on input. **WHY THIS PAGE MISSED IT:** qwen.ai/blog is client-rendered and returns nothing to plain fetches; the fix is to enumerate the Qwen org on Hugging Face by createdAt, which is how all three were found. **ALSO ON HF:** Qwen-Drive-1.0-4B (8/27), an autonomous-driving model outside this page's scope.

FLAGSHIP PREVIEW -- QWEN3.8-MAX (2026-07-19 at WAIC Shanghai, press-reported): Alibaba previewed **Qwen3.8-Max**, a **2.4-trillion-parameter sparse-MoE multimodal model** (text/image/video/docs), days after Moonshot's Kimi K3 open-weights launch. Alibaba positions it as 'second only to [Claude] Fable 5.' Access at preview is CLOSED: via Alibaba's Token Plan subscription plus the Qoder / QoderWork agentic platforms, at roughly 10% of standard pricing; **open weights 'promised soon' with no date or license, and no model card, benchmark table, or third-party eval yet** -- treat the ranking claim as unverified vendor positioning. SEPARATELY (2026-07-21) the Qwen team released **Qwen-Image-3.0**, a 3rd-gen image model (up to ~4,500-token prompts, legible text down to ~10px, 12 languages, UI/web + live-data rendering) -- but a transparency regression: **Qwen Chat access only, no open weights, no benchmarks, no technical report** (a departure from Qwen's prior Apache-2.0 same-day-weights image releases)

USER-CREATED AI AGENTS SHUT DOWN -- TOOK EFFECT 2026-07-15 (confirmed): Alibaba's Qwen app (and ByteDance's Doubao, simultaneously) **shut down user-created/custom AI agent features on July 15, 2026** as China's rules on 'humanlike AI interaction services' (the anthropomorphic-AI/companion rules) took effect that day; Qwen had already disabled humanlike interactive agents ~7/10 ahead of the broader cutoff. IMPORTANT DATA DETAIL: the platforms diverge -- **Doubao** offers read-only access to old agent data until **October 15, 2026** (unrecoverable after), but **Qwen offers no migration path at all**: agent configurations and conversation histories are permanently deleted. This is a consumer-app change -- the Qwen models/API are unaffected. Separately (SCMP, early July, press-reported): Alibaba cut Qwen3.7-Max by ~80% and Qwen3.7-Plus by ~60% for international users on the Qoder platform during off-peak hours (10pm-8am Beijing = US working hours) -- an explicit play for US developer demand amid a broader Chinese token price war

FLAGSHIP GA + VENDOR NUMBERS CONFIRMED (verified 2026-06-09): **Qwen 3.7 Max** shipped 2026-05-20 ('Qwen3.7: The Agent Frontier' on the official Qwen blog) -- the benchmark figures that circulated in aggregators are now vendor-published: **SWE-Bench Pro 60.6%, Terminal-Bench 2.0-Terminus 69.7%, GPQA 92.4%**, plus SWE-Multilingual 78.3%, SciCode 53.5%, QwenSVG 1608. PROPRIETARY / API-only on DashScope/Model Studio. Official pricing: **$2.50/M input, $7.50/M output list -- with a 50% launch promo ($1.25/$3.75) running until 2026-06-22**. Separately, **Qwen3.7-Plus** launched ~2026-06-02 on Bailian/Model Studio: multimodal (image + video input, text output) with GUI-agent, deep reasoning, and tool invocation at $0.40/$1.60 per 1M -- press-reported agent scores (ScreenSpot Pro 79.0, AndroidWorld 81.0) await vendor tables. Both models confirmed on Alibaba's official Model Studio model list