Qwen3 30B A3B
Alibaba's nimble MoE model that switches between deep reasoning and fast chat, with open weights and tool use.
Overview
What is Qwen3 30B A3B
Qwen3 30B A3B is an open-weights mixture-of-experts language model from Alibaba's Qwen team, released in April 2025. It activates roughly 3.3 billion parameters per token from a 30.5-billion-parameter pool, supports tool use and web search, and can toggle between thinking and non-thinking modes for reasoning or speed.
Running it privately on Venice
On Venice, Qwen3 30B A3B runs inside a TEE with end-to-end encryption and zero retention — your prompts are not stored or profiled. You get native tool use and web search capabilities on an open-weight, Apache 2.0 model without Big-Tech surveillance. It is an efficient MoE choice for private agents and long-context workflows.
Assessment
Strengths and limitations
- Open weights under Apache 2.0, enabling self-hosting, fine-tuning, and commercial reuse without restriction.
- Efficient MoE architecture: only ~3.3B parameters activate per token, cutting inference cost while maintaining strong reasoning.
- Seamless switching between thinking mode for complex math, coding, and logic, and non-thinking mode for fast, efficient dialogue.
- Strong tool-use and agent capabilities with native function calling and web search support on Venice.
- Multilingual coverage across 100+ languages and dialects with robust instruction following and translation.
- Competitive with larger models: officially reported to outperform QwQ-32B despite activating 10× fewer parameters.
- Smaller active parameter count (3.3B) limits peak reasoning depth compared to dense 32B+ or large frontier models.
- Text-only input: no vision or audio capabilities, so multimodal tasks require a separate model in the pipeline.
- MoE routing tables add memory overhead for local deployment even though active compute per token is low.
- Safety alignment is present; the model is not uncensored on Venice.
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 (Qwen Team)
- Released
- April 28, 2025
- Architecture
- Mixture-of-Experts (MoE), 48 layers, 128 experts (8 active)
- Parameters
- 30.5B total / 3.3B activated per token
- Open weights
- Yes — Apache 2.0
- Context window
- 256K tokens
- Max output
- 32.768K tokens
- Capabilities
- Function calling, Web search
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Mar 2026
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": "e2ee-qwen3-30b-a3b-p",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.19 per 1M input tokens and $0.69 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Context window | Open weights | Price (Venice) | Best for |
|---|---|---|---|---|
| Qwen3 30B A3B | 256K tokens | Yes | $0.19 in · $0.69 out / 1M | Tool use & agents |
| DeepSeek V3.2 | 160K tokens | Yes | $0.33 in · $0.48 out / 1M | Coding & MoE efficiency |
| Google Gemma 4 31B Instruct | 256K tokens | Yes | $0.12 in · $0.36 out / 1M | Budget long-context |
| Kimi K2.6 | 256K tokens | Yes | $0.75 in · $3.50 out / 1M | Premium reasoning |
Open-weights MoE that toggles thinking mode; best value for tool-using agents with a 256K context window.
Use cases
What it is good for
- 01Private AI agents that need tool calling and web search without surveillance or data retention.
- 02Cost-efficient coding assistance, mathematical reasoning, and step-by-step debugging.
- 03Multilingual customer support, translation, and content generation pipelines.
- 04Long-context document analysis and summarization using the 256K token window on Venice.
- 05Edge or self-hosted deployments where open weights and Apache 2.0 licensing are required.
Prompting
Getting better results
Enable thinking mode for step-by-step math, logic, or coding problems; disable it for quick Q&A and creative writing.
Provide explicit tool schemas when using function calling — the model follows structured instructions and JSON outputs reliably.
For multilingual tasks, specify the target dialect, tone, and domain directly in the prompt to improve alignment.
Version history
Initial open-weights MoE release with thinking/non-thinking modes and tool use.
FAQ
Frequently asked questions
Qwen3 30B A3B is an open-weights mixture-of-experts language model released by Alibaba's Qwen team in April 2025. It has 30.5 billion total parameters with about 3.3 billion activated per token, supports tool use and web search, and can switch between a thinking mode for deep reasoning and a non-thinking mode for fast conversation.
Venice charges $0.19 per 1 million input tokens and $0.69 per 1 million output tokens. There is no subscription required; you pay per token with credits.
You can try it free on Venice — new accounts include a free tier with no credit card required. Beyond the free limit, usage is billed per token at $0.19 in and $0.69 out per 1M tokens.
Yes. The model is released under the Apache 2.0 license, so weights are open and available for self-hosting, fine-tuning, or commercial use without restriction.
Yes. On Venice it supports function calling and web search, making it suitable for agentic workflows that need to query APIs or retrieve live information.
Qwen3 30B A3B offers a 256K context window and lower input pricing on Venice, while DeepSeek V3.2 is also an MoE with strong coding performance. Choose Qwen3 for tool-heavy, long-context agents at lower cost; choose DeepSeek V3.2 if you prefer its output pricing and coding focus.
Venice runs it in a private TEE with end-to-end encryption and zero retention — your prompts are not stored, profiled, or used for training. It is one of the most private ways to run an open-weights model.
Thinking mode activates deeper chain-of-thought reasoning for math, coding, and complex logic, while non-thinking mode skips that step for faster, more efficient general conversation. You can toggle between them per prompt.
Run Qwen3 30B A3B privately
No prompt logging. No data used for training.