Alibaba Qwen
Alibaba's Qwen open-weight models: Qwen3.8 runs from a 27B Apache 2.0 vision-language model to a 2.4T MoE, plus low-cost APIs in Model Studio
Overview
Qwen is the model family from Alibaba Cloud's Qwen team, and one of the largest sources of open-weight models: the Qwen organization on Hugging Face holds 469 models. The newest open generation, Qwen3.8 (August 2026), includes Qwen3.8-27B, a dense vision-language model that reads images and video, with a 262,144-token native context that extends to 1M tokens, under Apache 2.0; Qwen3.8-Flash-Next; and Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter mixture-of-experts "Max" model that Alibaba released with a custom license carrying commercial conditions (per SCMP and eWeek). The earlier Qwen3.6 line (35B-A3B and 27B) is also still listed. Not every Qwen model is open: Alibaba keeps API-only models such as qwen3.7-plus and qwen4-max-preview in its Model Studio service, where Qwen also sells hosted access to its open models at low per-token prices. Check the license file of each checkpoint before commercial use, because Qwen now mixes Apache 2.0 with custom licenses. Also from Alibaba: Qwen Chat, the consumer chat app.
The Verdict
Who Should Use Qwen?
Best For
- Self-hosting a strong mid-size model (Qwen3.8-27B on one high-end GPU)
- Long-context and vision tasks with open weights
- Chinese and multilingual products
- Teams that want the same model family both self-hosted and as a cheap API
Not Ideal For
- Buyers who need one simple license across the family
- Latency-sensitive local use with default settings (heavy reasoning output)
- Teams restricted from using models built by Chinese companies
What's Great
- Apache 2.0 on Qwen3.8-27B, with 262K native context extendable to 1M
- Huge catalog: 469 models from tiny to 2.4T parameters
- Qwen3.8-27B passed 7.2M Hugging Face downloads
- Low API prices: qwen3.8-flash at $0.15 in, $0.47 out per 1M (Singapore)
Watch Out For
- Qwen3.8-27B "wildly overthinks by default", using many reasoning tokens
- The Max-class Qwen3.8 model has a custom license with commercial restrictions
- Benchmarks are Alibaba-reported; independent reproduction lagged at launch
- The main GitHub repo has not been updated since January 2026
Pricing
View all features & details
Open Models
- Qwen3.8-27B: dense, vision-language, Apache 2.0, 262K native context (Aug 2026)
- Qwen3.8-Flash-Next (check its license file)
- Qwen3.8-2.4T-A95B: Max-class MoE, custom license
- Qwen3.6-35B-A3B and Qwen3.6-27B
- FP8 variants of most models
API Models (Model Studio)
- qwen4-max-preview, qwen3.7-plus, qwen3.8-flash (text)
- qwen3.8-omni-flash (omni), qwen-image-3.0-pro (image)
- qwen-audio-3.1, qwen3.7-text-embedding, qwen3-rerank
- qwen-max: $1.60 in, $6.40 out per 1M
- China (Beijing) region prices are lower than Singapore
Real-World Usage
Hugging Face Footprint (Nov 2026)
- 469 models in the Qwen organization
- Qwen3.8-27B: 7.23M downloads
- Qwen3.8-Flash-Next: 3.28M downloads
Community Reports
- One user ran Qwen3.8-27B at 215K context on a single consumer GPU at about 42 tokens a second
- Reviewers call it a strong open model but note slow, verbose reasoning
How It Compares
| Feature | Qwen3.8 (open) | Google Gemma 4 | Mistral Small 4 | NVIDIA Nemotron 3 |
|---|---|---|---|---|
| Largest open model | 2.4T MoE | 31B dense | 119B MoE | 550B MoE (Ultra) |
| Max context | 262K native, 1M extended | 256K | 256K | 1M |
| License | Apache 2.0 (27B), custom (Max) | Apache 2.0 | Apache 2.0 | NVIDIA or OpenMDW licenses |
| Hosted API by maker | Yes, Model Studio | AI Studio trial | Yes, La Plateforme | build.nvidia.com trial |
| Vision input | Yes | Yes | Yes | Yes (VL and Omni models) |
Sources: Qwen · Hugging Face · Qwen3.8-27B · Model Studio models · Model Studio pricing · eWeek (Aug 2026) · GitHub
Is this your tool? Add a badge