Qwen4 is coming to Runware
Alibaba's next generation of Qwen models, expected to build on Qwen3.8-Max with deeper knowledge, stronger coding and more capable agents. Coming soon to Runware's OpenAI-compatible endpoint.
- Expected API models
- Max
- Plus
- Flash
- Expected open-weight sizes
- 27B
- 4B
Built for real-world tasks.
Alibaba's announcement promises deeper knowledge, richer world understanding, and stronger coding and agentic capabilities.
- World knowledge
- Perception
- Coding
- Office productivity
- Front-end design
- Agents and GUI operations
What to expect from Qwen4
Built on Qwen3.8-Max
Alibaba says Qwen4 builds on Qwen3.8-Max's complex-task delivery, with stronger knowledge, perception and professional tool use.
A new architecture
Alibaba calls Qwen3.8-Flash-Next an experimental preview of the architecture behind Qwen4: 125B parameters with 6B active, plus n-gram embeddings that add capacity with less compute.
Knowledge and perception
Alibaba expects stronger world knowledge and perception, with a better understanding of complex environments and professional tasks.
Coding, office work and agents
Expected gains in coding, office productivity and front-end design, and in agent work, including GUI operations.
What we know so far
Alibaba hasn't published Qwen4's full specs yet. This is the expected lineup, plus what Alibaba's docs list for the Qwen4-Max preview, which may change at release.
- Developer
- Alibaba (Qwen team)
- API models
- Max, Plus, Flash
- Open-weight sizes
- 27B, 4B
- Context (Max preview)
- 1M tokens, 393K max output
- Input (Max preview)
- Text, images, video
- Reasoning (Max preview)
- On by default, low to xhigh
- Tools (Max preview)
- Function calling, JSON Schema output
- On Runware
- Coming soon
The Qwen3.8-Max baseline
Qwen4's benchmark results aren't published yet. Until they are, these are Alibaba's published scores for Qwen3.8-Max, the model it builds on.
- Qwen4-MaxComing soon
- GPT-5.6 Sol88.8%
- Qwen3.8-Max86.6%
- Fable 584.6%
- Opus 4.884.6%
- Qwen3.7-Max74.5%
- Qwen4-MaxComing soon
- Qwen3.8-Max93.0%
- GPT-5.6 Sol90.5%
- Fable 588.8%
- Opus 4.880.3%
- Qwen3.7-Max64.8%
- Qwen4-MaxComing soon
- Fable 584.5%
- Qwen3.8-Max75.1%
- GPT-5.6 Sol74.0%
- Opus 4.869.8%
- Qwen3.7-Max56.5%
- Qwen4-MaxComing soon
- GPT-5.6 Sol94.1%
- Qwen3.8-Max92.6%
- Fable 592.6%
- Qwen3.7-Max92.4%
- Opus 4.892.0%
| Benchmark | Qwen4-Max | Qwen3.8-Max | GPT-5.6 Sol | Fable 5 | Opus 4.8 | Qwen3.7-Max |
|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | Coming soon | 86.6% | 88.8% | 84.6% | 84.6% | 74.5% |
| PaperBench | Coming soon | 93.0% | 90.5% | 88.8% | 80.3% | 64.8% |
| AndroidBench | Coming soon | 75.1% | 74.0% | 84.5% | 69.8% | 56.5% |
| GPQA Diamond | Coming soon | 92.6% | 94.1% | 92.6% | 92.0% | 92.4% |
Alibaba-published figures from the Qwen3.8-Max model card. Fable 5 results may involve fallbacks.
LLMs on Runware today
The Runware API also hosts other state-of-the-art LLMs, all on the same endpoint and API key.
Frequently asked questions
What is Qwen4?
Alibaba's next generation of Qwen language models. Alibaba's preview lists three API models, Max, Plus and Flash, and open-weight models at 27B and 4B. It builds on Qwen3.8-Max, with expected gains in knowledge, perception, coding, office work and agents.
When will Qwen4 be available on Runware?
Qwen4 is coming soon, and there's no launch date yet. Leave your details below and we'll let you know as soon as it's live.
How will Qwen4 compare with Qwen3.8-Max?
Alibaba describes Qwen4 as building on Qwen3.8-Max's complex-task delivery, with stronger knowledge, perception and professional tool use. Qwen4's benchmark results aren't published yet, so the scores on this page are Qwen3.8-Max's.
Is Qwen4 multimodal?
Yes. Alibaba's docs for the Qwen4-Max preview list text, image and video input, for screenshot analysis, document understanding and video analysis, with text output. The preview's specs may change at release.
How much will Qwen4 cost on Runware?
Pricing is published at launch. Like every LLM on Runware, it will be billed per token, with no subscription.
How will I call Qwen4?
Through Runware's OpenAI-compatible endpoint, with the same API key and billing as every other model. If you already call an LLM on Runware, switching to Qwen4 means changing the model ID.
Can I use Qwen models on Runware today?
Alibaba's Qwen Image and Wan models are live on Runware for image and video generation. For language tasks, the LLMs listed on this page are live on the same endpoint now.
Get notified when Qwen4 launches
Leave your details and we'll let you know as soon as Qwen4 is live on Runware. We can also talk through rate limits and dedicated capacity.