Research

Alibaba Qwen Model Lineage and Roadmap 2026

Alibaba Qwen release history 2024-2026: Qwen 2, QwQ-32B, Qwen 2.5-Max, Qwen3 family (Apr 2025), Qwen3-VL, Qwen 3.5 (Feb 2026), Qwen 3.6 Plus, 3.6 Max Preview. Snapshot for 2026-05-15.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: May 2026

What this is

Alibaba's Qwen line is the most prolific open-weight family in 2026. Apache 2.0 across most releases, MoE architecture since late 2024, and an aggressive cadence that has put Qwen at or near the top of every open-weight benchmark in 2026. This page is a 2026-05-15 release-by-release reference.

Qwen Release Timeline

DateReleaseNotable
Late 2024Qwen 2.0MoE; first to outperform Llama 3 on MMLU/HumanEval
Nov 2024QwQ-32B-Previewo1-style reasoning model, Apache 2.0
Dec 25, 2024QvQ-72B-PreviewVisual reasoning research model
Jan 29, 2025Qwen 2.5-MaxFlagship-tier release
Jan 2025Qwen 2.5-VL (3/7/32/72B)Multimodal lineup
Apr 28, 2025Qwen 3 familyDense 0.6B-32B + MoE 30B-A3B, 235B-A22B, Apache 2.0
Sep 23, 2025Qwen3-VL-235B (Instruct + Thinking)Top open multimodal benchmark
Late 2025Qwen3-Coder-PlusCoding agent series on Qwen3
Feb 16, 2026Qwen 3.5 + Qwen 3.5-PlusOpen-weight + proprietary tiers
Apr 2026Qwen 3.5-Omni + Qwen 3.6-PlusProprietary; multimodal omni model
Apr 2026Qwen 3.6-Max PreviewStronger world knowledge + agentic coding

Family at a Glance (Mid-2026)

FamilyArchitectureLicenseBest at
Qwen 3 (dense)0.6B-32B denseApache 2.0On-device + edge inference
Qwen 3 MoE30B-A3B, 235B-A22BApache 2.0Server inference at frontier-near quality
Qwen3-CoderCoding-tuned Qwen3Apache 2.0Coding agents, SWE-Bench class
Qwen3-VLVision + textApache 2.0Multimodal benchmarks (SOTA open)
Qwen 3.5 (open) + 3.5-Plus (proprietary)MoEMixedQ1 2026 flagship
Qwen 3.6-Plus + 3.6-Max PreviewProprietary MoEClosedAgentic coding + world knowledge

Six Things the Lineage Tells You

  1. Qwen ships faster than any non-Chinese open-weight family. 11+ flagship releases in 18 months.
  2. Apache 2.0 across the open tier. No MAU clauses, no field-of-use restrictions — the friendliest open licence among major Chinese labs.
  3. Dense + MoE coverage. 0.6B dense for edge through 235B-A22B MoE for server is the broadest size range from any single lab.
  4. Modality split: open for base + coder + vision; proprietary for top-tier flagship. Qwen 3.5-Plus and 3.6-Plus stay closed; the open tier still leads the open-weight leaderboard.
  5. Qwen3-VL 235B leads the open multimodal benchmarks. Roughly comparable to closed frontier multimodal on most evals.
  6. Agentic coding is the 2026 product focus. Qwen3-Coder + 3.6-Max Preview both lean into agentic / multi-step coding workloads.

What This Means for AI Visibility

Qwen is the open-weight family most likely to be deployed by enterprises, developers, and local-AI projects. Brand visibility on Qwen-served surfaces (Cursor with Qwen routing, on-prem deployments, OpenClaw with Qwen backend) matters in 2026, and the underlying model uses different default sources than ChatGPT or Claude. Test brand presence on Qwen separately from US-frontier surfaces.

Methodology

Release dates and capabilities combine the Wikipedia Qwen article, the official Qwen blog, QwenLM GitHub, Qwen on HuggingFace, and Alibaba Cloud Model Studio documentation.

How Presenc AI Helps

Presenc AI tracks brand visibility across Qwen-served surfaces (cloud Qwen, on-prem Qwen3, OpenClaw with Qwen backend, Qwen-routed agents) so brand teams see how the open-weight default mix changes their AI mention footprint over time.

Frequently Asked Questions

Qwen 3.6-Plus and Qwen 3.6-Max Preview, released April 2026, are the latest proprietary tier. Qwen 3.5 (Apache 2.0, February 2026) is the latest open-weight flagship. Qwen3-VL-235B-A22B is the latest open multimodal flagship.
The core Qwen 3 and Qwen 3.5 families ship under Apache 2.0 with no MAU clauses. The Plus and Max-Preview variants of Qwen 3.5 and Qwen 3.6 are proprietary. Alibaba's pattern is to release open base models and reserve flagship-Plus tiers as proprietary.
Qwen 3 MoE matches or slightly leads Llama 4 Maverick on most non-coding benchmarks, especially MMLU-Pro and GPQA Diamond. Llama 4 Scout leads on context window (10M tokens vs Qwen's typical 128K-256K). License-wise Apache 2.0 (Qwen) is friendlier than the Meta custom license (Llama 4).
A coding-tuned Qwen 3 variant designed for agentic coding workflows. Late-2025 releases (Qwen3-Coder-Plus) compete with DeepSeek-Coder and proprietary coding models on SWE-Bench and HumanEval. Open-weight under Apache 2.0.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.