Qwen family (Qwen3.8)
Open-weight model line from Alibaba (Qwen team). Verified from Hugging Face cards on 2026-09-30 and re-checked 2026-10-02; Qwen3.8 releases in Aug 2026.
Current models (Hugging Face Qwen)
| Model | Size | Licence | Notes |
|---|---|---|---|
| Qwen3.8-2.4T-A95B (the open weights of Qwen3.8-Max) | 2.4T total / 95B active | custom qwen3.8-max | 92 layers, 512 experts (10 routed + 1 shared), Gated DeltaNet + Gated Attention; 262,144 native context, extensible to ~1.01M; card: SWE-bench Pro 67.7, GPQA Diamond 92.6 (vendor) |
| Qwen3.8-27B | 27B dense, vision encoder | Apache 2.0 | 262K native, to 1M; text, image, video |
| Qwen3.8-Flash-Next | 125B total / 6B active (+51B n-gram embedding) | qwen-community-1.0 | 262K native, to 1M |
Hugging Face shows the Max weights updated Aug 12 and the 27B on Aug 14. The commercial Qwen3.8-Max variant (with vision) is served via Qwen Cloud.
Licence nuance
The Max licence is not Apache. Per its text (summarised from a fetch): products above 100M monthly active users or 50M+ revenue over 12 months need a separate licence; commercial derivatives are restricted for high-revenue businesses; “AI Work Assistant” products are specifically restricted. Read the licence file yourself before commercial use.
Generation timeline
| Generation | Date | Open weights? |
|---|---|---|
| Qwen3.5 | 2026-02-16 | Yes, plus proprietary Qwen3.5-Plus |
| Qwen3.6 | April 2026 (35B-A3B 04-16, 27B 04-22) | Yes, Apache 2.0; Qwen3.6-Plus proprietary |
| Qwen3.7 | Max 2026-05-20, Plus 2026-06-01 | No: proprietary only; no open 3.7 shipped |
| Qwen3.8 | August 2026 (Max weights 08-12, 27B 08-14, Flash-Next updated 08-27 on HF) | Yes (licences above) |
| Qwen 4 | previewed 2026-09-22 at Apsara (tiers Max, Flash, Plus, 27B) | Not released; no date, pricing or specs published |
Dates for 3.5-3.7 and the Qwen 4 preview come from Wikipedia and secondary trackers (Qwen3.8 rows are from the Hugging Face cards). The Max open-weight model omits cloud-only features such as image input and a non-thinking mode (secondary report). The name Qwen is also written Tongyi Qianwen. Hub notes: qwen (company/resource note), qwen-3-dot-8-max, company note Alibaba. This note also absorbed the former qwen-models-overview (2025 notes), 2026-10-02.
Earlier generations (2025 to early 2026; vendor and video-creator claims, attributed)
- Qwen2.5-Max (Jan 2025): large MoE trained on 20T+ tokens; Qwen’s blog (qwenlm.github.io/blog/qwen2.5-max) reports it ahead of GPT-4o and DeepSeek-V3 on benchmarks such as Arena-Hard (89.4) and LiveBench (62.2); vendor figures, not independently re-checked. Qwen2.5-VL (vision-language, PC/phone control per TechCrunch 2025-01-27) and Qwen2.5-Omni (text/image/video/audio in, text/audio out) came from the same generation.
- Qwen3 (April 2025): dense models 0.6B-32B and MoE 30B-A3B / 235B-A22B, hybrid thinking/non-thinking mode, 119 languages, ~36T training tokens, Apache 2.0, open weights on Hugging Face, ModelScope and Ollama (TechCrunch 2025-04-28; Qwen GitHub). Qwen3-Max (Sept 2025) was the larger proprietary flagship.
- QwQ: the earlier reasoning model, see qwq.
- Qwen3.6 (April 2026) as covered on YouTube: AI Stack Engineer (2026-04-23) reported Qwen 3.6 Max leading several coding benchmarks including SWE-bench and Terminal-bench and beating Claude Opus 4.6 on an agentic coding benchmark (creator claim); Income Stream Surfers (2026-04-08) used Qwen 3.6 Plus with OpenCode as a free Claude Code alternative. Free chat access via chat.qwen.ai and API via Alibaba Cloud Model Studio and OpenRouter were mentioned; mid-2026 free-tier reductions were reported in the same coverage, not verified here.
- Deployment frameworks named by the Qwen3 README: SGLang, vLLM, llama.cpp, Ollama, MLX.
Unverified points
- Qwen 4 preview and Qwen3.5-3.7 dates rest on Wikipedia and secondary trackers (accessed 2026-10-02); no Alibaba primary page was fetched.
- Qwen3.8-Max licence thresholds are from a summarised fetch of the licence text; read the file before commercial use.
- Flash-Next size: the card says 125B / 6B active plus 51B n-gram embedding and 4B MTP; the Hugging Face listing shows 180B in total.
Sources
- https://huggingface.co/Qwen (accessed 2026-09-30)
- https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B and its LICENSE (accessed 2026-09-30)
- https://huggingface.co/Qwen/Qwen3.8-27B (accessed 2026-09-30)
- https://huggingface.co/Qwen/Qwen3.8-Flash-Next (accessed 2026-09-30)
- https://huggingface.co/Qwen (re-checked 2026-10-02)
- https://en.wikipedia.org/wiki/Qwen (accessed 2026-10-02)
- https://github.com/QwenLM/Qwen3.8 (listed in search, 2026-10-02)
- Secondary: https://codersera.com/blog/qwen-3-5-complete-guide-2026/ ; https://www.versely.studio/blog/qwen-4-announced-at-apsara-2026 (accessed 2026-10-02)
Related
- qwen-3-dot-8-max (per-model note), qwq (reasoning model), qwen (company/resource note)
- https://qwenlm.github.io/blog/qwen2.5-max/ ; https://github.com/QwenLM/Qwen3 ; https://techcrunch.com/2025/04/28/alibaba-unveils-qwen-3-a-family-of-hybrid-ai-reasoning-models/ ; https://techcrunch.com/2025/01/27/alibabas-qwen-team-releases-ai-models-that-can-control-pcs-and-phones/
- YouTube: AI Stack Engineer, 2026-04-23 (rreYfRDZftY); Income Stream Surfers, 2026-04-08 (-38eEugIwok); AI Bites, 2025-05-01 (L5-eLxU2tb8)