Model record
Qwen3.6 35B-A3B
Alibaba Cloud · Apache 2.0 · verified
What this model is
Qwen reports 35B total parameters and 3B activated parameters, with 8 routed experts plus 1 shared expert per token.
The architecture is mixture of experts. Memory must hold the full weight set, while active parameters affect the speed heuristic. Experts can be offloaded to system memory, but that trades speed for capacity.
Memory and quantization
Decimal GB. Sizes combine the published GGUF weight file with the F16 multimodal projector required for image and video input.
| Quantization | Published files | 8K estimate |
|---|---|---|
| MXFP4 | 22.61 GB | 24.27 GB |
| Q4_K_M | 23.03 GB | 24.70 GB |
| Q8_0 | 37.80 GB | 39.47 GB |
Estimate formula: published weight files + 0.1678 GB per 8K tokens × context / 8K + 1.5 GB runtime overhead. Sizes are decimal GB, not GiB. At 8K context, the smallest listed build is about 24.3 GB before OS and application reserve.
Full-attention KV estimate from the public config. Ten of 40 layers use full attention; linear-attention state is small and excluded from this conservative per-token figure.
KV method source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json
Speed heuristic
Not benchmarked here. The calculator uses 3B active parameters only as a relative speed heuristic.
Active path: 3B parameters. This is a relative speed signal only; it is not a tokens/sec benchmark.
If the full weights do not fit in accelerator memory but do fit in system memory, the calculator marks the result as a tight CPU-offload fit. Expect substantially slower generation and verify the runtime's MoE offload support.
Runtimes
- Ollama
- llama.cpp
- MLX
- vLLM
ollama run qwen3.6:35b-a3bLicense restrictions
Apache 2.0 permits commercial use and modification; retain notices and comply with the patent and termination terms.
This is a plain-language summary, not legal advice. Read the linked license before commercial deployment or redistribution.
Known issues
- The native context is 262K tokens; long prompts can exhaust memory even when the quantized weights fit.
- The GGUF package needs the multimodal projector for vision and video input.
Sources
- https://huggingface.co/Qwen/Qwen3.6-35B-A3B
- https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF
- https://ollama.com/library/qwen3.6
- https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json
This page is an independent compatibility guide, not official documentation. Verify the primary source before downloading or deploying a model.