OnyX-9B
Agentic model for terminal and tool use
2 min read
OnyX-9B by Ajelix is a 9.47-billion-parameter agentic language model built on Qwen3.5-9B. It is trained for multi-step tool and terminal sessions, including reasoning, function calling, response analysis, and retries. The model retains the ability to answer direct questions concisely and supports a native context of up to 262k tokens. It is useful for developers and engineers for automating terminal tasks, debugging code, and working with repositories.
Key features
- Based on the Qwen3.5-9B architecture, retaining the built-in vision encoder.
- Trained for interaction via bash command calls and the generation of reasoning markup.
- Supports native tool use via the standard Qwen chat template.
- Distributed as a merged checkpoint for Transformers and vLLM, as well as a standalone PEFT adapter (LoRA, a small add-on layer over the base model).
What you can do
- Run locally via Transformers for text queries and script execution.
- Deploy the model backend using vLLM for high-throughput generation.
- Use the PEFT adapter on top of the base Qwen/Qwen3.5-9B model.
- Integrate the model into system administration and debugging automation pipelines.
Example
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "AjelixAI/Onyx-9B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")
tools = [{
"type": "function",
"function": {
"name": "bash",
"description": "Run shell commands in the container and return terminal output.",
"parameters": {
"type": "object",
"properties": {
"command": {
"type": "string"
},
"required": ["command"],
},
},
},
}]
messages = [{"role": "user", "content": "List the files in /app."}]
text = tokenizer.apply_chat_template(
messages, tools=tools, tokenize=False, add_generation_prompt=True, enable_thinking=True
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Read more
DetailsBenchmarks: official results have not been published yet.
Hardware requirements: the merged checkpoint weighs 18.8 GB (bfloat16 support required).
Limitations: trained primarily on the bash tool; the safety of executing destructive commands has not been tested.
License: Apache-2.0.










Comments