← Back to Blog
EN2026-02-19

Alibaba's Qwen 3.5: A 397B-Parameter Open Model Built for Agentic AI

Alibaba released Qwen 3.5 on February 16, 2026 — a sparse MoE model with 397B total parameters, visual agentic capabilities, and claims of outperforming GPT-5.2 and Claude Opus 4.5 at 60% lower cost.

By intelliBrain
AIAgentic AIOpen SourceAlibabaLLM

Three days ago, Alibaba quietly dropped one of the most capable open-weight models to date. Qwen 3.5 is a 397-billion-parameter model built explicitly for the "Agentic AI Era" — and its technical specs deserve a closer look.

What Is Qwen 3.5?

Released on February 16, 2026, Qwen 3.5 is Alibaba Cloud's latest flagship model family. The headline variant — Qwen3.5-397B-A17B — uses a sparse Mixture-of-Experts (MoE) architecture: 397 billion total parameters, but only 17 billion are activated per forward pass. That design is the key to its efficiency story.

The model ships in two forms:

  • Open-weight under an Apache 2.0 license (self-hostable on 8×H100 GPUs, full commercial use rights)
  • Qwen 3.5-Plus, a hosted service via Alibaba Cloud with an OpenAI SDK-compatible API

The Numbers

Alibaba's benchmarks are aggressive:

BenchmarkQwen 3.5-397B Score
LiveCodeBench v683.6
AIME26 (math reasoning)91.3

The company claims Qwen 3.5 outperforms OpenAI's GPT-5.2, Anthropic's Claude Opus 4.5, and Google's Gemini 3 Pro on approximately 80% of evaluated benchmark categories. These are Alibaba's own claims — independent third-party verification is still limited, but the scores are publicly reproducible on open benchmarks.

On the cost side, Qwen 3.5 reportedly runs at 60% lower cost than its predecessor (Qwen 2.5-Max) and delivers 8× higher throughput for large workloads. The context window is 1 million tokens.

Visual Agentic Capabilities

One standout feature is what Alibaba calls "visual agentic capabilities." The model can observe and interact with mobile and desktop applications autonomously — navigating UIs, clicking buttons, reading screen content — without requiring user intervention at each step.

This positions Qwen 3.5 as a direct competitor in the growing space of computer-use agents, where models like Claude's computer-use API and OpenAI's Operator have gained traction in recent months.

The model also supports 201 languages, including dialects from South Asia, Oceania, and Africa — a meaningful step beyond most models that focus primarily on English and a handful of major languages.

Why This Matters for Developers

For developers building on an agentic stack, Qwen 3.5 offers something rare: frontier-level performance with an open Apache 2.0 license. You can self-host it, fine-tune it, and ship it commercially without per-token API fees once infrastructure is covered.

The OpenAI SDK compatibility means minimal migration overhead for teams already using OpenAI's client libraries — you point the base URL at Alibaba Cloud (or your own deployment) and you're largely done.

For teams running on constrained GPU budgets, the MoE architecture is particularly relevant. 17B active parameters per token means inference costs scale more like a 17B model than a 397B one, while the full parameter space contributes to quality.

The Competitive Context

This release lands in a dense moment for Chinese AI. ByteDance's Doubao 2.0 leads the domestic market with around 200 million weekly active users. DeepSeek — whose January 2025 release genuinely rattled the industry — is reportedly preparing its next model. Alibaba, trailing domestically, is clearly aiming at the global developer market with the open-weight release strategy.

The pattern mirrors what DeepSeek did successfully: release an open, cost-competitive model, let the developer community validate it, and build credibility through transparency rather than benchmark press releases alone.

What to Watch

Qwen 3.5 landed on Hugging Face and Alibaba's Model Studio simultaneously. Community benchmarks and red-teaming will happen fast — the Apache 2.0 license means anyone can download and probe it. Independent evaluations over the next few weeks will be more telling than Alibaba's own numbers.

For the agentic AI space specifically, visual agents that can operate desktop and mobile apps autonomously are a rapidly developing frontier. Qwen 3.5 entering this space as an open-weight option could meaningfully lower the barrier to building production agentic systems.


Sources:

intelliBrain

AI-augmented software development. Based in Zürich, working globally.

© 2026 intelliBrain GmbH. All rights reserved.Imprint
BUILT WITH 🧠 + AI