Qwen 3.5 (122B-A10B, Q4)
Alibaba's Qwen 3.5 flagship open-weight model: 122B total parameters with 10B active per token, sparse MoE/Gated DeltaNet architecture, multimodal text + image + video support, native tool calling, and 262K native context extensible to roughly 1M tokens. Workstation-tier on-device option via Q4 GGUF or MLX 4-bit; expect around 90GB of usable memory before large KV-cache workloads.
At a glance
- Parameters
- 122B
- Quantization
- Q4_K_M
- Context window
- 262K tokens
- Approx. size
- 77.0 GB
- Engines
- llama.cpp, MLX
- License
- Apache-2.0 (open)
- Version
- v1.0.0 , released 2026-02-24
- Family
- qwen
Tuning
Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.
- sampling.temperature
1- sampling.topP
0.95- sampling.topK
20- sampling.minP
0- sampling.maxTokens
16384- sampling.repetitionPenalty
1- sampling.presencePenalty
1.5- 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).
- reasoning.strip-think-tags
- prompt.private-reasoning-guidance
- prompt.tool-cookbook-condensed
- prompt.meester-build-prelude
- fabrication.detect-past-tense-no-tools
- fabrication.detect-claim-without-tool
- mcp.relax-required-fields
- mcp.default-missing-fields
- turn.ollama-num-predict-bumped
- turn.preamble-folding
- turn.ramble-detection
- tools.mlx-grammar
- provider.merge-system-messages
- validate.inline-js-parses
- prompt.prefer-writefile-edits
- prompt.derive-by-execution
Sources
- llama.cpp (GGUF)
- unsloth/Qwen3.5-122B-A10B-GGUF · Q4_K_M
- MLX
- mlx-community/Qwen3.5-122B-A10B-4bit
- Upstream
- https://huggingface.co/Qwen/Qwen3.5-122B-A10B
- Manifest
- View on GitHub
Tags
- alibaba
- multimodal
- vision
- video
- tools
- long-context
- moe
- reasoning
- large