Orbit

Expert-Specialized Fine-Tune (ESFT) of Qwen/Qwen3.5-35B-A3B for Vietnamese multi-turn tool-use, trained on reasoning traces.

Architecture is unchanged from the base model.

This model was training on the distillation dataset from claude-opus-5, effort xhigh for my custome workflow

Capabilities

Multi-turn tool use

This model is trained to maintain context across multiple tool interactions, rather than treating each tool call as an isolated operation. This makes it suitable for workflows where the result of one action determines the next action.

Tool selection

The model is trained on trajectories containing tool selection and execution, allowing it to reason about:

• which tool should be used
• when a tool call is necessary
• what arguments should be provided
• how to interpret tool results
• whether additional actions are required

The training data contains high-effort reasoning trajectorie

Training

• ESFT: only selected MoE experts are trained (router frozen).
• trainable: ~0.94B of 35.6B parameters (2.6%)
• expert selection: top_p = 0.20, ~7.5 of 256 experts per layer
• single NVIDIA GB10 (121 GB unified memory)

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tk = AutoTokenizer.from_pretrained("beyoru/Clawd-Agent")
model = AutoModelForCausalLM.from_pretrained("beyoru/Clawd-Agent", dtype="auto", device_map="auto")

msgs = [{"role": "user", "content": "..."}]
ids = tk.apply_chat_template(msgs, tools=TOOLS, add_generation_prompt=True, return_tensors="pt")

Serving with vLLM:

vllm serve beyoru/Orbit --max-model-len 8192

The chat template emits <think>\n in the generation prompt, so the model continues the reasoning block and closes it with </think> before the answer.

Note

Inherits the base model's MIT license. Fine-tuned on a narrow task distribution — evaluate on your own workload before relying on it for anything outside multi-turn tool use.