Qwen3.6-35B-A3B is an open-weight, multimodal Mixture-of-Experts (MoE) foundation model developed by Alibaba Cloud’s Qwen team. Featuring 35 billion total parameters with only 3 billion activated per token, it combines a hybrid Gated DeltaNet linear attention and MoE architecture with native multimodal understanding (text, image, video), a 262K native context window (extensible to 1M+ via YaRN), and cross-turn thinking preservation for agentic coding and reasoning workflows—released under the permissive Apache 2.0 license.

Running long-context multimodal reasoning and agentic tool workflows on open-source infrastructure traditionally required heavy compute footprints and led to slow inference speeds. Qwen3.6-35B-A3B addresses this by pairing an active parameter footprint of just 3B with a high-capacity 35B routing structure, allowing developers to achieve frontier-level coding, multi-turn tool calling, and multimodal reasoning on single-GPU enterprise or local setups.

Architecture & Technical Profile at a Glance

Feature / Metric	Qwen3.6-35B-A3B Specification	Core Developer & Production Benefit
Model Architecture	35B Total / 3B Active (MoE)	256 total experts (8 routed + 1 shared per token) for high-efficiency inference.
Attention Mechanism	Hybrid Gated DeltaNet + Gated Attention	Substantially lowers memory overhead and speeds up long-sequence decoding.
Context Length	262,144 Tokens (Native)	Extensible up to 1.01M tokens via YaRN RoPE scaling for full repo ingestion.
Modalities	Multimodal (Text, Image, Video)	Unified visual-language understanding across documents, screenshots, and video clips.
Thinking Architecture	Preserved Reasoning Traces (preserve_thinking)	Keeps chain-of-thought visible and accessible across multi-turn agent conversations.
Serving & Precision	BF16, FP8, NVFP4 on vLLM & SGLang	Fits onto a single H100/H200 or MI300X in FP8, and Blackwell in NVFP4.
License	Apache 2.0	Unrestricted commercial deployment, local hosting, and fine-tuning.

Core Breakthroughs & Capabilities

• Active-3B Mixture of Experts Efficiency: By activating only 3 billion parameters per token out of 35 billion total, the model achieves inference throughput exceeding 120+ tokens/second while preserving the broad knowledge base and reasoning depth of a much larger dense network.

• Preserved Thinking for Agentic Pipelines: Unlike typical reasoning models that wipe their chain-of-thought scratchpad between prompt turns, Qwen3.6 supports multi-turn thinking preservation. Agents can review why prior decisions or tool calls were made, making multi-step coding and execution workflows significantly more reliable.

• Native Multimodal Understanding: Processes text, high-resolution images, and video files in the same forward pass. It easily handles visual UI debugging, document OCR, chart reasoning, and frame-by-frame video question answering.

• Turnkey Open-Source Deployment: Fully supported across major inference engines (vLLM, SGLang, Ollama, MLX) with first-class Hugging Face and ModelScope checkpoints, allowing teams to deploy locally or integrate drop-in OpenAI-compatible endpoints into agents like Hermes Agent, OpenHands, and Cursor.