AlibabaAlibaba·💬 Text Generation

Qwen 3.5 397B

ReasoningVisionCodeFunction CallingWeb Searchanonymized
🧠 Try in Intelligence →Try on Venice.ai ↗
Quick reference
Qwen 3.5 397B — TLDR
  • 🆕 Alibaba's Qwen3.5-generation flagship, released February 2026.
  • 🧠 397B-parameter Mixture-of-Experts activating only 17B per token.
  • 👁️ Native vision-language model trained on multimodal data.
  • 🔧 Default thinking mode, tool calling and MCP integration.
  • 📏 128K-token context window for long inputs.
  • 🎯 Targets reasoning, coding and agentic workflows.
  • 📚 Artificial Analysis scores it 45 on its Intelligence Index.
  • 🔒 Open weights under the Apache-2.0 license.
💰 Pricing
$0.750 / $4.50
per 1M · input / output
📏 Context
128K tokens
📅 On Venice since
Feb 16, 2026
107 days ago
Provider

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Read full profile →
46 models on Venice
17 text · 16 video · 5 image · 4 inpaint · 2 embedding · 2 tts
Since Jan 11, 2025

About this model

Qwen 3.5 397B (model ID Qwen3.5-397B-A17B) is the flagship of Alibaba's Qwen3.5 generation, released February 16, 2026 under an Apache-2.0 license. It is a sparse Mixture-of-Experts system with 397B total parameters but only 17B activated per token, giving a wide expert pool at relatively low per-token compute.

Compared with its same-family predecessor Qwen 3 235B A22B Instruct 2507, Qwen3.5 nearly doubles the total parameter count while activating fewer parameters per token (17B versus 22B), trading a larger, sparser routing pool for cheaper inference. It is also among the first Qwen flagship releases with native vision input, trained on multimodal data rather than bolting a vision encoder on afterward.

The model targets reasoning, coding, multilingual tasks and agentic workflows, with a default thinking mode that emits internal reasoning before answering, plus tool calling and Model Context Protocol support. It serves a 128K-token context window for long documents and codebases. Independent evaluator Artificial Analysis places it at 45 on its Intelligence Index. As an open-weight release distributed under Apache-2.0, it can be self-hosted, fine-tuned and deployed without usage restrictions, continuing Alibaba's pattern of shipping its larger Qwen models with permissive licensing.

This About section is AI-generated from public sources (Claude Opus 4.8), with no human editing. It may contain inaccuracies — verify critical details against the sources listed above.

Data sources: Venice API · HuggingFace · Wikipedia — enrichment updated 1d ago