---

license: apache-2.0

base_model:

• Qwen/Qwen3.6-35B-A3B

tags:

• moe
• expert-pruning
• agentic
• reasoning
• gguf

pipeline_tag: text-generation

---

Devlin Alpha 22B A3B

Devlin Alpha 22B A3B is a development version of Devlin Mini, a lightweight model focused on agentic tasks based on Qwen3.6 35B A3B. This version of the model is intended only for testing purposes and is not recommended for production. Benchmarks pending.

Table of Contents 📝

1. ▶ Usage
2. 📃 License
3. 🙏 Acknowledgements

<a name="usage"/>

▶ Usage

1. Download Models

Download models using huggingface-cli:

pip install "huggingface_hub[cli]"
huggingface-cli download Devlin-AI/Devlin-Alpha-22B-A3B-GGUF --local-dir ./Devlin-Alpha-22B-A3B-GGUF

You can also download directly from this page.

2. Inference

To use these GGUF files, you'll need a compatible inference engine like llama.cpp or clients built on top of it (e.g., Ollama, LM Studio, KoboldCpp, text-generation-webui with a llama.cpp backend).

> ⚠️ Important: Always pass --jinja when loading with llama.cpp so the Qwen3.6 chat template is applied correctly. Without it, the model may emit malformed turns.

> Note: Text-only, and there's no MTP head, so draft/MTP speculative-decoding flags don't apply.

llama.cpp (server)

llama-server -hf Devlin-AI/Devlin-Alpha-22B-A3B-GGUF --port 8000 -c 262144 --jinja

Recommended sampling parameters (thinking mode): temperature=1.0, top_p=0.95, top_k=20, with --chat-template-kwargs "{\"enable_thinking\":true,\"preserve_thinking\":true}". For precise coding tasks: temperature=0.6, top_p=0.95, top_k=20. For non-thinking mode: temperature=0.7, top_p=0.8, top_k=20, presence_penalty=1.5, with --chat-template-kwargs "{\"enable_thinking\":false}".

Parsing Reasoning Traces

Responses begin with a <think> … </think> block containing the chain-of-thought, followed by the final answer. To split them:

if "</think>" in text:
    reasoning, answer = text.split("</think>", 1)
    reasoning = reasoning.replace("<think>", "").strip()
    answer = answer.strip()
else:
    reasoning, answer = "", text.strip()

Ollama

ollama run hf.co/Devlin-AI/Devlin-Alpha-22B-A3B-GGUF

<a name="license"/>

📃 License

This model is a derivative work of Qwen/Qwen3.6-35B-A3B, licensed under the Apache 2.0 License, and is therefore distributed under the terms of the same license.

<a name="acknowledgements"/>

🙏 Acknowledgements

• Qwen Team for the base model:

* Qwen3.6-35B-A3B on Hugging Face

• The llama.cpp project and its contributors for the GGUF format and the incredible tooling that makes local LLM inference accessible.

* llama.cpp GitHub Repository