Overview
- Tool name: Qwen3.5
- Developer: Alibaba (Tongyi Lab)
- Official website: https://qwen.ai
- Access: free download on ModelScope / HuggingFace, or the Alibaba Cloud Bailian API
- Category: AI Models
- License: Apache 2.0 (open releases)
Qwen3.5 was open-sourced on February 16, 2026 (Chinese New Year's Eve) and immediately topped the open-source leaderboard. Highlights:
- Natively multimodal: first pretrained on mixed vision+text tokens (Qwen3 was text-only), lifting reasoning, coding and agentic scores across the board
- Small beats big: the flagship open release Qwen3.5-397B-A17B has 397B total parameters with only 17B active, yet outperforms the trillion-parameter Qwen3-Max; 60% less VRAM to deploy and up to 19x inference throughput
- Extreme API value: Qwen3.5-Plus from CNY 0.8 per million tokens - roughly 1/18 of Gemini 3 Pro
Official benchmarks: MMLU-Pro 87.8 (above GPT-5.2), GPQA Diamond 88.4 (above Claude 4.5), IFBench 76.5 (record); BFCL-V4 and BrowseComp agent evaluations beat Gemini 3 Pro and GPT-5.2. Several authoritative vision-understanding evaluations rank it best, and it ingests up to 2 hours of video (1M-token context) directly.
The family spans Plus (API flagship), 397B-A17B (open flagship) and 0.8B/2B/4B/9B small models (open-sourced in March), plus later Qwen3.6 edge variants and the API-only Qwen3.7-Max.
API pricing (official, Alibaba Cloud Bailian)
- Qwen3.5-Plus: from CNY 0.8 per million tokens (official launch figure; check the Bailian console for live rates)
- Open weights are free to download; self-hosting costs compute only
Pricing verified as of October 2026 (official launch communications).
Who it is for
- Agents and tool calling: top-tier BFCL-V4 / BrowseComp scores with stable multi-turn tool chains
- Long-video understanding: 1M-token context ingests 2-hour videos directly - ideal for analysis and summarization
- Cost-efficient production: flagship capability at CNY 0.8 per million tokens crushes high-volume costs
- Private deployment: Apache 2.0 with sizes from 0.8B to 397B; Day-0 support from NVIDIA, AMD, Apple, Ascend and more
Limitations
- Multimodal output (image generation) still relies on the dedicated Qwen-Image line; Qwen3.5 focuses on understanding
- Access paths differ by region; commercial use in China generally goes through Bailian
- The Qwen family iterates fast (3.5 / 3.6 / 3.7-Max sold in parallel) - match the tier to the task
Frequently asked questions
Is Qwen3.5 the strongest open model?
At launch in February 2026 Alibaba claimed the top open-source spot (MMLU-Pro 87.8, GPQA 88.4); it sits in the same first tier as DeepSeek V4, GLM-5 and Kimi K3.
What does the Qwen3.5 API cost?
Qwen3.5-Plus from CNY 0.8 per million tokens (official figure), about 1/18 of Gemini 3 Pro; live rates on Alibaba Cloud Bailian.
Is it open source?
Yes - Qwen3.5-397B-A17B and the small models are Apache 2.0 and free for commercial use; only Qwen3.7-Max stays API-only.
Qwen3.5 or DeepSeek V4?
Both are MIT/Apache first-tier open models: choose Qwen3.5 for native multimodal (vision/video) and the full size family; choose DeepSeek V4 for 1M-context pure-text value (Flash at $0.14/$0.28).
*This page is not affiliated with Alibaba. Benchmarks are vendor-reported; pricing per Alibaba Cloud Bailian. Verified as of October 2026.*
