Qwen 3.6 (35B-A3B, Q4)
Alibaba's Qwen 3.6 at 35B total parameters with only 3B active per token (Mixture-of-Experts: 256 experts, 8 routed + 1 shared). Multimodal (text + image + video), 256K native context (extensible to ~1M with YaRN), explicit native tool-calling support. MoE sparsity keeps inference light despite the larger total parameter count.
At a glance
- Parameters
- 35B
- Quantization
- Q4_K_M
- Context window
- 262K tokens
- Approx. size
- 21.2 GB
- Engines
- llama.cpp, MLX
- License
- Apache-2.0 (open)
- Version
- v1.0.0 , released 2026-06-02
- Family
- qwen
Tuning
Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.
- sampling.temperature
0.7- sampling.topP
0.8- sampling.topK
20- sampling.minP
0- sampling.maxTokens
8192- reasoning.thinkingBudget
4096- reasoning.enableThinking
true- profiles
- thinking-general, thinking-coding, thinking-precise, instruct, creative
Model behaviors
Client-side behavior modules the catalog enables for this model (reasoning-tag handling, fabrication detection, prompt shaping).
- provider.merge-system-messages
- reasoning.strip-think-tags
- mcp.compact-tool-schemas
- prompt.private-reasoning-guidance
- prompt.tool-cookbook-condensed
- fabrication.detect-past-tense-no-tools
- turn.ollama-num-predict-bumped
- turn.preamble-folding
- turn.ramble-detection
- tools.mlx-grammar
Sources
- llama.cpp (GGUF)
- lmstudio-community/Qwen3.6-35B-A3B-GGUF ·
Qwen3.6-35B-A3B-Q4_K_M.gguf· Q4_K_M - MLX
- mlx-community/Qwen3.6-35B-A3B-4bit
- Upstream
- https://huggingface.co/Qwen/Qwen3.6-35B-A3B
- Manifest
- View on GitHub
Tags
- alibaba
- multimodal
- vision
- video
- tools
- long-context
- moe