LLMAnonymized

Qwen 3.5 397B

Alibaba's flagship open-weight multimodal model with 397B parameters, native vision, and efficient MoE architecture — now on Venice.

Maker
Alibaba Cloud
Modality
Text + vision
Context
128K tokens
Open weights
Yes — Apache 2.0

Overview

What is Qwen 3.5 397B

Qwen 3.5 397B is Alibaba's flagship open-weight multimodal AI, released in February 2026. Built on a sparse Mixture-of-Experts architecture with 397 billion total parameters and 17 billion active per token, it supports vision, reasoning, tool use, and code generation, delivering high efficiency and broad language coverage across 201 languages.

Running it privately on Venice

On Venice, Qwen 3.5 397B runs with anonymized privacy — your prompts are never stored or profiled. You get full access to its multimodal and agentic capabilities, including vision and web search, without sacrificing sovereignty. The model is uncensored in operation, enabling permissionless use for developers and enterprises.

AnonymizedNo prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • Native multimodal support: unified vision-language processing via early fusion training, enabling image and video input.
  • Efficient inference: sparse MoE architecture activates only 17B of 397B parameters per token, balancing performance and cost.
  • Strong agentic and reasoning capabilities, with official support for tool calling, web search, and structured JSON output.
  • Extensive language coverage: supports 201 languages, making it one of the most globally accessible open models.
  • Open weights and commercially permissive license (Apache 2.0) enable self-hosting, fine-tuning, and enterprise integration.
Limitations
  • Not the most cost-efficient option: higher output cost compared to rivals like DeepSeek V4 Flash.
  • Closed reasoning mode in some deployments; the full 'thinking' capability may require specific configuration.
  • While open weights, the training data and full pipeline are not fully transparent.

Capabilities

What it supports

  • Tool use / function calling
  • Vision (image input)
  • Reasoning
  • Web search
  • Code-optimized
  • Structured output (JSON schema)
  • Audio input
  • Video input
  • Multiple image inputs
  • Log probabilities

Specifications

Datasheet

Maker
Alibaba Cloud
Released
February 16, 2026
Architecture
Sparse Mixture-of-Experts (MoE)
Parameters
397B total / 17B active per token
Open weights
Yes — Apache 2.0
Context window
128K tokens
Max output
32.768K tokens
Capabilities
Vision, Function calling, Reasoning, Web search, Code-optimized
Privacy on Venice
Anonymized — prompts not stored
Available on Venice since
Feb 2026
License
Apache 2.0

API

Call it from your code

Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-5-397b-a17b",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $0.75 per 1M input tokens and $4.50 per 1M output tokens.

Input / 1M tokens
$0.75
Per 1M tokens
Output / 1M tokens
$4.50
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelMax outputStrongest atOpen weightsPrice (Venice)
Qwen 3.5 397B32.768K tokensMultimodal agents, code, reasoningYes$0.75 in · $4.50 out / 1M
Claude Opus 532.768K tokensComplex reasoning, enterprise tasksNo$6 in · $30 out / 1M
DeepSeek V4 Flash 073132.768K tokensSpeed and cost efficiencyNo$0.17 in · $0.35 out / 1M
Google Gemma 4 31B Instruct32.768K tokensEfficiency and open accessYes$0.12 in · $0.36 out / 1M

Open, efficient MoE model with strong vision and agentic performance.

Use cases

What it is good for

  1. 01Multimodal RAG applications combining text and image inputs.
  2. 02Internationalized agents and chatbots requiring broad language support.
  3. 03Code generation and reasoning workflows with tool integration.
  4. 04Enterprise automation leveraging vision, web search, and function calling.
  5. 05Privacy-sensitive deployments where prompt retention is prohibited.

Prompting

Getting better results

Use explicit image URLs or base64-encoded images to trigger vision mode.

Enable tool use by specifying available functions in JSON schema format.

For complex reasoning tasks, prompt with 'think step by step' to engage reasoning mode if available.

Leverage its multilingual strength by specifying non-English languages directly in the prompt.

Version history

Qwen3 235B A22B
2025-07

Previous generation, text-only

Qwen3.5 397B A17B
2026-02

Current — native vision, MoE, open weights

FAQ

Frequently asked questions

Qwen 3.5 397B is Alibaba's flagship open-weight multimodal AI model, released in February 2026. It features a 397B-parameter sparse Mixture-of-Experts architecture with 17B active per token, supporting vision, reasoning, tool use, and code generation across 201 languages.

Yes. Qwen 3.5 397B is open weights under the Apache 2.0 license, allowing free commercial use, self-hosting, and fine-tuning. The model weights are publicly available on Hugging Face.

On Venice, it costs $0.75 per 1M input tokens and $4.50 per 1M output tokens. There are no upfront fees or subscriptions — billing is per token used.

Yes. It natively supports image and video input through unified vision-language training, making it one of the first open models to integrate multimodal understanding directly into its core architecture.

Yes. The model supports function calling and tool use, including structured JSON output and integration with external tools like web search and code execution environments.

Qwen 3.5 397B is open, multimodal, and far more cost-effective, while Claude Opus 5 excels in complex reasoning and enterprise readiness but is closed and significantly more expensive. Choose Qwen for open, vision-capable agents; Claude for closed-loop enterprise workflows.

The model has a 128K token context window on Venice, with research versions supporting up to 262K tokens via YaRN extension. The maximum output length is 32,768 tokens.

Run Qwen 3.5 397B privately

No prompt logging. No data used for training.