Qwen 3.6 35B A3B FP8
Alibaba's open-weights 35B-parameter MoE with 3B active per token, built for agentic coding, reasoning, and tool use.
Overview
What is Qwen 3.6 35B A3B FP8
Qwen 3.6 35B A3B FP8 is Alibaba's open-weights Mixture-of-Experts language model with 35 billion total and 3 billion active parameters. Released in April 2026 under Apache 2.0, it specializes in agentic coding, chain-of-thought reasoning, and tool use, and is quantized to FP8 for efficient, near-original inference.
Running it privately on Venice
On Venice, Qwen 3.6 35B A3B FP8 runs inside a TEE with end-to-end encryption and zero retention — your prompts and code are never stored or profiled. You get full access to its reasoning, tool use, and web-search capabilities at $0.18/1M input tokens, with the sovereignty of open weights and the privacy of a private inference stack.
Assessment
Strengths and limitations
- Open-weights Apache 2.0 model with efficient MoE inference (3B active out of 35B total).
- Strong agentic coding skills: handles frontend workflows, repository-level reasoning, and iterative debugging via Thinking Preservation.
- Native tool use, reasoning, and web search support on Venice.
- FP8 quantization delivers near-original performance with reduced memory footprint.
- Runs privately in a TEE with end-to-end encryption and zero retention on Venice.
- 32K context window is shorter than many open-weight rivals (160K–256K).
- Vision encoder present in weights is not exposed through Venice's text endpoint.
- Benchmark aggregates trail top-tier closed frontier models on some agentic tasks.
- FP8 quantization, while efficient, may introduce minor precision trade-offs for sensitive numerical workloads.
- Not uncensored: retains standard safety alignment.
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 2026
- Architecture
- Mixture-of-Experts (Gated DeltaNet + Gated Attention)
- Parameters
- 35B total / 3B active
- Layers
- 40
- Hidden dimension
- 2048
- Experts
- 256 (8 routed + 1 shared)
- Context window
- 32K tokens
- Max output
- 4.096K tokens
- Capabilities
- Function calling, Reasoning, Web search, Code-optimized
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- May 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-qwen3-6-35b-a3b",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.18 per 1M input tokens and $1.18 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 | Strongest at | Open weights | Price (Venice) |
|---|---|---|---|---|
| Qwen 3.6 35B A3B FP8 | 32K tokens | Agentic coding & reasoning | Yes | $0.18 in · $1.18 out / 1M |
| DeepSeek V3.2 | 160K tokens | General reasoning & long context | Yes | $0.33 in · $0.48 out / 1M |
| Google Gemma 4 31B Instruct | 256K tokens | Long-context lightweight tasks | Yes | $0.12 in · $0.36 out / 1M |
| Claude Sonnet 4.6 | 1M tokens | Agentic tasks & safety | No | $3.60 in · $18 out / 1M |
Top open-weight pick for agentic coding with 3B active parameters and native tool support.
Use cases
What it is good for
- 01Agentic software development and repository-level coding.
- 02Iterative debugging with reasoning context preservation.
- 03Tool-augmented automation (function calling, web search).
- 04Private analysis of sensitive code or data under zero-retention policy.
- 05Efficient open-weights inference run privately on Venice's stack.
Prompting
Getting better results
Enable reasoning mode for complex logic or multi-step coding problems.
Use function calling to let the model orchestrate tools and external APIs.
Keep prompts under 32K tokens; summarize long files before inclusion.
Iterate conversations to leverage Thinking Preservation for context-aware refinements.
Be explicit about repository structure when asking for codebase-wide changes.
Version history
Preceding series referenced in the Qwen3.6 release.
Current open-weight release with FP8 quantization and agentic coding upgrades.
FAQ
Frequently asked questions
It is Alibaba's open-weights Mixture-of-Experts model with 35 billion total and 3 billion active parameters, released in April 2026. It is optimized for agentic coding, reasoning, and tool use, and is quantized to FP8 for efficient inference.
Pricing is $0.18 per 1 million input tokens and $1.18 per 1 million output tokens. Cached input tokens are billed at $0.06 per 1 million.
Yes. The model is released under the Apache 2.0 license and its weights are openly available on Hugging Face, allowing self-hosting and fine-tuning.
Yes. On Venice it supports function calling, reasoning, web search, and code-optimized generation, making it suitable for agentic workflows.
Choose Qwen if you want a 3B-active-parameter MoE focused on agentic coding with very low input pricing. Choose DeepSeek V3.2 if you need a 160K context window and cheaper output tokens.
No. While it is open-weights, it retains standard safety alignment and is not marketed as an uncensored model. Venice runs it as-is.
The model supports a 32K token context window and up to 4.096K tokens of max output per request on Venice.
Yes. On Venice it runs in a TEE with end-to-end encryption and zero retention, meaning your prompts are not stored, profiled, or used for training.
'A3B' stands for 3 billion activated parameters per forward pass. The model has 35 billion total parameters but only activates 3 billion at a time for efficient inference.
Run Qwen 3.6 35B A3B FP8 privately
No prompt logging. No data used for training.