🔥 New: Qwen-Image-2.1 Open-Sourced (Sep 20, 2026)
Alibaba Qwen open-sourced Qwen-Image-2.1 on September 20, 2026 - a unified text-to-image generation and image editing model whose visual component packs just 7B parameters (32 Single-Stream DiT layers). It natively supports generating and editing transparent (RGBA) images, accepts up to 10 reference images for multi-subject composition (group portraits, virtual try-on, interior design), and handles local edits via circles, painted annotations, or separate masks - with improved fidelity for both people and products. On the official Qwen-Image-Bench it scores 60.28, ranking #1 among open-source image models, ahead of Nano Banana 2.0 (59.82) and GPT Image 1.5 (59.65). Open weights are live on Hugging Face, ModelScope, and GitHub with day-zero support in ComfyUI, Diffusers, vLLM-Omni, and SGLang; it runs on 24GB consumer GPUs. Note the license: Qwen Research License - free for research and non-commercial use; commercial deployment requires a separate agreement from Qwen. Read the official release →
Latest Update (September 2026)
Latest (Sep 23): Qwen officially launched the Qwen-Audio-3.1 voice model family - five models spanning ASR, TTS and Realtime voice interaction, plus next-generation ASR-Next and TTS-Next. Prices are down across the line: TTS ~70%, Realtime ~85%, and ASR up to 95% cheaper; Qwen-Audio-3.1-ASR-Flash is just RMB 0.8 in / 2.7 out per 1M tokens, and TTS-Flash is RMB 1.5 in / 12 out. The new ASR model beat open-source rivals in 6 of 11 subsets on public dialect benchmarks (KeSpeech, WSYue) with an average character error rate of only 4.55%. Most APIs are already live on Alibaba Cloud Model Studio (Bailian), further lowering the cost of building voice agents.
Latest (Sep 22): At Apsara Conference, Alibaba confirmed a leadership change for Qwen - researcher Liu Daheng formally takes over as Qwen LLM project lead (ATH business group, Token Foundry) and delivered a keynote titled "Qwen: Toward Real-World Agents". He previously oversaw post-training and the Coding team, so the full Qwen training chain now sits with one lead. Qwen3.8-Max ranks #7 on Artificial Analysis' Intelligence Index and #1 on its Agentic Index, while the open-source ecosystem leads Hugging Face with 3 billion downloads, 300k derivative models and 460+ open models. Next-gen Qwen4 is already in training, targeting 5-10 trillion parameters.
Latest (Sep 22): The Qwen AI platform was upgraded around "putting agents into production", adding Agent services and industry AI solutions on top of model services, plus an API Express Mode (1.5-2x TPS, switch model ID to activate), Agent Studio and the Qwen AI Cockpit. FlashBoot/UniScheduler cut cold-start from 1200s to 70s (10k pods in 1 minute), first-token latency drops 38%, and prompt caching saves up to 95%. Alibaba Cloud MaaS customers grew 6x in the past year; Alibaba expects 20GW+ of cloud compute by 2032.
Beyond text models, the Qwen3 voice ecosystem is heating up: Alibaba Cloud Model Studio (Bailian) now serves official Qwen3-TTS-Flash and Qwen3-ASR-Flash-Realtime endpoints (17 expressive voices, multi-language, emotion recognition, dialect support). Independent inference providers are also pushing the open-weight Qwen3-TTS 1.7B model to the top of Coval's voice AI benchmarks - #1 latency for speech-to-text and #1 word-error-rate for text-to-speech as of mid-September 2026. For developers building voice agents, Qwen voice APIs combine leading open-source quality with low per-character cost, and you can get started with the same free token allowance.
Qwen3 Model Pricing Overview
Alibaba Cloud offers a range of Qwen3 models through DashScope/Model Studio. Prices below are per 1 million tokens, in USD for international access and CNY for China mainland.
| Model | Input (per M tokens) | Output (per M tokens) | Context | Free Tier |
|---|---|---|---|---|
| Qwen Code CLI | FREE | FREE | 262K | 1,000-2,000 calls/day |
| Qwen3.8-Max (flagship) | ~$1.65 (¥12) | ~$4.95 (¥36) | 128K | 70M tokens (new users) |
| Qwen3.7-Max | ~$0.83 (¥6)* | ~$2.48 (¥18)* | 128K | 70M tokens (new users) |
| Qwen3-Plus | ~$0.55 (¥4) | ~$1.65 (¥12) | 128K | Free quota available |
| Qwen3.7-Flash (0-32K ctx) | $0.03 (¥0.225) | $0.13 (¥0.974) | 256K | Free quota available |
| Qwen3.7-Flash (32-256K ctx) | $0.10-$0.21 | $0.41-$0.83 | 256K | Free quota available |
| Qwen3-Coder-Plus | ~$0.83 (¥6) | ~$2.48 (¥18) | 262K | Via Qwen Code free |
| Qwen3-Coder 480B (Vertex AI) | $0.22 | $1.80 | 262K | Via Qwen Code free |
| Qwen3.5-Omni (multimodal) | ~$0.25 (¥1.8) | ~$2.17 (¥15.8) | 128K | 1M tokens free |
* Qwen3.7-Max prices shown are after the 50% limited-time discount (standard: ¥12/¥36). USD conversions approximate at 7.25 CNY/USD.
Free Tier Breakdown
Option 1: Qwen Code CLI (Always Free)
- Mainland China: 2,000 API calls/day via Qwen OAuth or ModelScope
- International: 1,000 API calls/day via OpenRouter
- No token limits within each call
- 60 requests/minute rate limit
- Uses Qwen3-Coder 480B model
npx @qwen-code/qwen-code@latestOption 2: New User Tokens
- 70 million+ free tokens across Qwen models
- Valid for 180 days (extended from 90 days)
- 100 AI image generation credits
- 50 seconds of video generation
- 200 CNY no-threshold coupon
- No credit card required (international) / real-name verification (China)
Option 3: OpenRouter Free Models
- qwen/qwen3-coder:free — 20 RPM, 200 requests/day
- qwen3-32b — 30 RPM, 1,000 requests/day on GitHub Models
API Quick Start
from openai import OpenAI
# International endpoint
client = OpenAI(
api_key="your-dashscope-key",
base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
)
response = client.chat.completions.create(
model="qwen3.7-max",
messages=[{"role": "user", "content": "Explain async/await in Python"}]
)
print(response.choices[0].message.content)Cost Comparison with Competitors
| Model | Input/M tokens | Output/M tokens | Free Tier |
|---|---|---|---|
| Qwen3.7-Flash | $0.03 | $0.13 | Yes (70M tokens) |
| DeepSeek V3 | $0.14 | $0.56 | No |
| GPT-4o Mini | $0.15 | $0.60 | Limited |
| Claude Haiku | $0.80 | $4.00 | No |
| Gemini Flash | $0.075 | $0.30 | Limited |
| Qwen Code (CLI) | FREE | FREE | 1,000-2,000 calls/day |
Money-Saving Tips
- Use Flash models for simple tasks: Qwen3.7-Flash is 90% cheaper than Max and sufficient for classification, summarization, and simple chatbots.
- Maximize free tier first: 70M tokens can handle approximately 70,000-100,000 typical API requests before needing to pay.
- Use Qwen Code for development: 2,000 free coding calls per day replaces $10-20/month Copilot subscriptions.
- Context caching discount: Qwen3.7-Flash offers reduced pricing for cached tokens (¥0.225/M vs ¥0.749/M for 32K context).
- Coding Plan Lite: Heavy coders can get unlimited Qwen3-Coder for 7.9 CNY (~$1.10) first month.
- Qoder paid plans from $20/month: for AI agent deployments, Qoder Pro bundles 2,000 Credits with concurrent agents.
Start Building for Free
Claim your 70M+ free tokens and $200 cloud credits today.
Get $200 Free Credits →No credit card required · 180-day AI token validity
→ Full Qwen Code setup guide · Free cloud hosting comparison