Qwen 3.5

Summary

Qwen 3.5 is Alibaba Cloud's open-weight multimodal model family, first released February 16, 2026 with a 397B-A17B MoE flagship and 201-language coverage. Nine variants shipped in 16 days, plus a proprietary Qwen 3.5-Omni adding audio and real-time interaction in April 2026.

Overview

Qwen 3.5 is Alibaba Cloud's open-weight model family, with the first variants released February 16, 2026 — timed to the eve of the Lunar New Year — as the centerpiece of Alibaba's response to the global open-weight AI race. The Qwen team subsequently completed an unusually rapid rollout of nine Qwen 3.5 family models in 16 days (releasing the Qwen 3.5 Small Model Series on March 2, 2026), and added Qwen 3.5-Omni — a native multimodal model spanning text, audio, video, and real-time interaction — in April 2026.

Qwen 3.5 is positioned by Alibaba as China's flagship contribution to the agentic AI race. The first release includes a 397B-parameter mixture-of-experts model with 17B activated parameters per token (397B-A17B), supporting native multimodal input across text, images, and video and fluency in 201 languages and dialects (up from 82 in the previous generation).

Specifications

  • Developer: Alibaba Cloud / Qwen team
  • Initial Release Date: February 16, 2026 (Qwen 3.5 + Qwen 3.5-Plus); Small Model Series March 2, 2026; Qwen 3.5-Omni April 2026
  • Type: Multimodal LLM family; includes MoE flagship and small/medium variants
  • Flagship Variant: 397B parameters total / 17B activated per token (397B-A17B MoE)
  • Language Coverage: 201 languages and dialects (up from 82 in prior generation)
  • License: Open weights for most variants; Qwen 3.5-Omni released April 2026 as proprietary
  • Modalities: Native multimodal (text + image + video) in main models; Qwen 3.5-Omni adds audio and real-time interaction
  • Access: Hugging Face, Alibaba Cloud Model Studio, GitHub (Qwen3.6 repo houses the 3.5 family per the Qwen team's repository convention)

Capabilities

Native Multimodal: Text, image, and video understanding within a single architecture. Qwen 3.5-Omni extends this to audio and real-time interaction (April 2026 proprietary release).

Multilingual Reach: 201 languages and dialects — among the broadest language coverage of any frontier-class open-weight model.

Agentic Focus: Alibaba's positioning emphasizes agentic AI deployment for enterprises. The flagship 397B-A17B MoE is sized for high-quality enterprise agentic workflows; the Small Model Series (released March 2) targets on-device and resource-constrained deployment.

Rapid Iteration: Nine Qwen 3.5 family models released in 16 days — one of the fastest open-weight rollouts on record, reflecting the maturity of Alibaba's training and release infrastructure.

Commercial Open-Weight Default: Most Qwen 3.5 variants are released as open weights, continuing the pattern that has made Qwen one of the most-downloaded open-weight model families on Hugging Face.

Limitations

Subsequent Family Versioning: The Qwen team's GitHub repository structure (referenced as "Qwen3.6" in the public GitHub naming convention) suggests rapid versioning is ongoing — production deployments should pin to specific model strings to avoid drift.

Geopolitical Considerations: Like other Chinese-origin frontier models, Qwen 3.5 use in U.S. enterprise environments often involves additional review around training data provenance and export-control posture. The model's use in restricted environments varies by sector and is evolving.

Qwen 3.5-Omni Closed: While the main Qwen 3.5 family is open-weight, the multimodal Omni variant (with audio and real-time interaction) was released in April 2026 as proprietary — splitting the family across distribution models.

Performance Benchmarks Vary: As with any model family spanning E2B-class on-device variants up to 397B-parameter MoE flagships, benchmark performance is highly variant-dependent and language-dependent. Validation on the specific deployment target is required.

Recent Developments

  • February 16, 2026: Initial Qwen 3.5 and Qwen 3.5-Plus release on Lunar New Year eve — with the 397B-A17B MoE as the flagship.
  • March 2, 2026: Qwen 3.5 Small Model Series released, completing nine Qwen 3.5 family models in 16 days.
  • April 2026: Qwen 3.5-Omni released as proprietary — native multimodal across text, audio, video, with real-time interaction support.
  • Industry Context: Qwen 3.5's release sits alongside DeepSeek V4 (April 24), Mistral Medium 3.5 (April 29), Gemma 4 (April 2), and Meta's Muse Spark (April 8) — making the open-weight tier in early-to-mid 2026 the most competitive it has ever been.
  • Distribution Strategy: Continues the open-weight default that has made Qwen one of the most-downloaded model families globally.

Last Updated

May 7, 2026

→ Back to Models