Qwen 3.8 Max
Alibaba's 2.4T MoE flagship — vision, reasoning, and code-optimized with 1M context and private execution on Venice.
Overview
What is Qwen 3.8 Max
Qwen 3.8 Max is Alibaba's flagship 2.4-trillion-parameter mixture-of-experts (MoE) multimodal model, released in August 2026. It supports vision, function calling, reasoning, web search, and code optimization, with a 1M-token context window and up to 131K output tokens, designed for agentic and enterprise workloads.
Running it privately on Venice
On Venice, Qwen 3.8 Max runs with anonymized privacy — your prompts are never stored or used for training. This enables permissionless, sovereign access to a top-tier proprietary model while maintaining enterprise-grade data hygiene. You get full tool use, vision, and long-context reasoning without surveillance.
Assessment
Strengths and limitations
- Top-tier performance on agentic benchmarks like OSWorld-Verified (86.1) and PaperBench, surpassing Fable 5 and GPT-5.6 Sol Max in reported results.
- Massive 1M-token context window and 131K max output, ideal for long-horizon tasks, code generation, and document processing.
- Multimodal input support: accepts images, video, and text with strong vision and reasoning integration.
- Highly cost-efficient for a model of its class, especially for output-heavy workloads.
- Optimized for code, tool use, and autonomous agent workflows with reliable function calling and web search.
- Not open-source or self-hostable: weights are proprietary and not currently available for on-premise deployment.
- Benchmark results depend heavily on extended time budgets (up to 12 hours), which may not reflect real-time use cases.
- No end-to-end encryption or TEE execution on Venice — privacy is anonymized but not fully sovereign at the hardware level.
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
- Released
- August 3, 2026
- Architecture
- Mixture-of-Experts (MoE)
- Parameters
- 2.4 trillion (active count undisclosed)
- Open weights
- No
- Context window
- 1,000K tokens
- Max output
- 131.072K tokens
- Capabilities
- Vision, Function calling, Reasoning, Web search, Code-optimized
- Privacy on Venice
- Anonymized — prompts not stored
- Available on Venice since
- Jul 2026
- License
- Proprietary
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": "qwen-3-8-max",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $2.50 per 1M input tokens and $7.50 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 | Best for | Max output | Vision | Open weights | Price (Venice) |
|---|---|---|---|---|---|
| Qwen 3.8 Max | Alibaba's flagship MoE model with top-tier agentic performance and 1M context. | 131K | Yes | No | $2.50 in · $7.50 out / 1M |
| Claude Fable 5 | Higher cost but strong in long-horizon reasoning; Qwen 3.8 Max undercuts it on price and benchmarks. | 128K | Yes | No | $12 in · $60 out / 1M |
| Claude Opus 5 | Competitive reasoning model, but Qwen 3.8 Max offers better price-performance. | 128K | Yes | No | $6 in · $30 out / 1M |
| DeepSeek V4 Flash 0731 | Much cheaper, but lacks vision and multimodal capabilities; different use case. | 128K | No | No | $0.17 in · $0.35 out / 1M |
Alibaba's flagship MoE model with top-tier agentic performance and 1M context.
Use cases
What it is good for
- 01Autonomous software engineering agents that use OS-level tools and code editors.
- 02Long-document analysis, summarization, and cross-reference tasks using the 1M context.
- 03Multimodal research workflows combining vision, web search, and reasoning.
- 04Enterprise automation pipelines requiring function calling and code generation.
- 05High-fidelity agent simulations with video and image inputs.
Prompting
Getting better results
Use explicit step-by-step reasoning in prompts to leverage its strong reasoning mode.
Include images or screenshots when asking for visual analysis — it handles multiple inputs well.
For code tasks, specify file paths and expected outputs to improve tool use accuracy.
Break long tasks into phases to stay within 131K output limits while maximizing context use.
Use web search when up-to-date information is critical — it integrates results effectively.
Version history
Predecessor model with smaller context and lower benchmark scores.
Current flagship — 2.4T MoE, 1M context, vision, and code optimization.
FAQ
Frequently asked questions
Qwen 3.8 Max is Alibaba's flagship 2.4-trillion-parameter mixture-of-experts (MoE) multimodal AI model, released in August 2026. It supports vision, function calling, reasoning, web search, and code generation with a 1M-token context window, designed for advanced agent and enterprise workflows.
On Venice, Qwen 3.8 Max costs $2.50 per 1M input tokens and $7.50 per 1M output tokens. Cached input is available at $0.31 per 1M tokens, making repeated access more efficient.
No. Qwen 3.8 Max is a proprietary model developed by Alibaba. While open weights are expected to be released in the near future, the model is currently closed and not available for self-hosting.
Yes. Qwen 3.8 Max supports image and video input, enabling multimodal reasoning and analysis. It can process multiple images in a single prompt and integrate visual data into its responses.
Yes. Qwen 3.8 Max supports function calling and tool use, allowing it to interact with external systems, execute code, and perform actions as part of an autonomous agent workflow.
Qwen 3.8 Max has a 1,000,000-token context window, one of the largest available, enabling it to process extremely long documents, codebases, or conversation histories in a single session.
Qwen 3.8 Max outperforms Claude Fable 5 on benchmarks like OSWorld-Verified and PaperBench, while costing significantly less per token. However, Fable 5 may still have advantages in certain reasoning tasks. Qwen offers superior price-performance for most agentic and code-heavy use cases.
Yes. On Venice, Qwen 3.8 Max runs under an anonymized privacy tier — your prompts are not stored, profiled, or used for training. This ensures private, permissionless access to the model without surveillance.
Run Qwen 3.8 Max privately
No prompt logging. No data used for training.