Qwen 3.8 2.4T
Alibaba's open-weight, 2.4T parameter sparse MoE flagship — strong in coding, research, and long-context reasoning with 262K token context on Venice.
Overview
What is Qwen 3.8 2.4T
Qwen 3.8 2.4T is Alibaba's flagship open-weight language model, released in August 2026. Built on a sparse Mixture-of-Experts architecture with 2.4 trillion total parameters and 95 billion active per token, it excels in coding, research, and long-horizon agentic tasks with a 262,144-token context window and support for reasoning control and web search.
Running it privately on Venice
On Venice, Qwen 3.8 2.4T runs with full privacy — your prompts are never stored, profiled, or used for training. This uncensored, open-weights model is available without surveillance, giving developers and researchers sovereign control over sensitive workflows. You get the full power of Alibaba’s latest open flagship, now with zero retention and permissionless access.
Assessment
Strengths and limitations
- Exceptional performance in coding, research, and long-horizon agentic tasks, with strong results on benchmarks like PaperBench (93.0) and OSWorld (86.1).
- Supports function calling, web search, and code-optimized generation, making it ideal for developer workflows and autonomous agents.
- Open weights allow for self-hosting, fine-tuning, and full transparency — a rare trait among flagship models.
- Large 262K context window with reasoning depth control via `reasoning_effort`, enabling complex, multi-step task execution.
- Text-only input: does not support vision or multimodal inputs despite the multimodal Qwen3.8-Max variant.
- Always runs in reasoning mode: cannot disable chain-of-thought, which may increase latency and cost for simple queries.
- Lower GPQA Diamond score (92.6) compared to GPT-5.6 Sol (94.1), indicating a slight gap in factual reasoning.
Capabilities
What it supports
- Tool use / function calling
- Reasoning
- Web search
- Code-optimized generation
- Vision input
- Video input
- Non-thinking mode
- Fine-tuning support
Specifications
Datasheet
- Maker
- Alibaba Cloud
- Open weights
- Yes — under custom qwen3.8-max license
- License
- Custom: qwen3.8-max
- Modes
- Reasoning modes: low, medium, xhigh; always reasons
- Context window
- 262.144K tokens
- Prompt length
- 262.144K tokens
- Released
- August 2026
- Architecture
- Sparse Mixture-of-Experts (MoE)
- Parameters
- 2.4T total, ~95B activated per step
- Max output
- 65.536K tokens
- Capabilities
- Function calling, Reasoning, Web search, Code-optimized
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Aug 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": "qwen-3-8-2-4t-a95b",
"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 | Context | Open weights | Price (Venice) |
|---|---|---|---|---|
| Qwen 3.8 2.4T | Open flagship coding & research | 262K tokens | Yes | $2.50 in · $7.50 out / 1M |
| Claude Fable 5 | Proprietary reasoning | 1M tokens | No | $12 in · $60 out / 1M |
| DeepSeek V4 Flash 0731 | Speed & cost | 1M tokens | Yes | $0.17 in · $0.35 out / 1M |
| GLM 5.1 | Balanced open model | 200K tokens | Yes | $1.10 in · $4.15 out / 1M |
Choose Qwen 3.8 2.4T when you need the highest reasoning and coding performance among open, private models — ideal for developers and researchers who demand transparency and control.
Use cases
What it is good for
- 01Long-form code generation and debugging with access to external tools and documentation.
- 02Research summarization and synthesis across large document sets within the 262K context window.
- 03Autonomous agent workflows requiring reliable multi-step planning and environment interaction.
- 04High-stakes reasoning tasks where transparency and auditability are required via open weights.
Prompting
Getting better results
Use `reasoning_effort=low` for faster, cheaper responses on simple queries.
Set `preserve_thinking=true` to retain reasoning context across conversation turns.
Include explicit tool calls in your prompt if the model doesn’t auto-invoke them.
Structure complex tasks as step-by-step instructions to leverage its agentic strengths.
Use web search mode for up-to-date factual queries outside its training data.
Keep prompts under 262K tokens to maximize output space for long-form responses.
Version history
Architectural foundation
Stability and utility improvements
Predecessor flagship
Current — open flagship with 2.4T MoE
FAQ
Frequently asked questions
Qwen 3.8 2.4T is Alibaba's flagship open-weight language model, released in August 2026. It features a sparse Mixture-of-Experts architecture with 2.4 trillion parameters and excels in coding, research, and long-context reasoning tasks with a 262K-token context window.
The model weights are open and available under a custom license called 'qwen3.8-max', allowing self-hosting and inspection. However, it is not free to run — on Venice, it costs $2.50 per million input tokens and $7.50 per million output tokens.
The model supports a context window of 262,144 tokens, with up to 65,536 tokens for output. This allows for extremely long conversations and document processing.
No. The Qwen 3.8 2.4T model available on Venice is text-only. While the hosted Qwen3.8-Max variant supports multimodal inputs, this open checkpoint does not.
Yes. The model supports function calling and tool use, making it suitable for agentic workflows, code execution, and integration with external APIs.
Qwen 3.8 2.4T is open and private, while Claude Fable 5 is closed and more expensive. Fable 5 leads in SWE-bench and HLE, but Qwen offers superior cost efficiency and sovereignty for developers who need open models.
The model is open weights and uncensored in the sense that it does not enforce policy-based filtering by default. However, deployment platforms may apply their own content policies.
Run Qwen 3.8 2.4T privately
No prompt logging. No data used for training.