A LoRA fine-tune of Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled trained on OpenClaw tool-call data — optimized for agentic reasoning with structured tool invocation.

The base model is a Claude 4.6 Opus reasoning distillation of Qwen3.5-27B. This fine-tune adds structured tool-calling capability in the OpenClaw XML format, making it suitable for local agentic deployments.

GGUF Downloads

Quantization	Size	Use case
Q4_K_M	15.4 GB	✅ Recommended — 24GB VRAM or 32GB unified memory
Q8_0	26.6 GB	Near-lossless, 32GB+ VRAM or 48GB unified memory

Training Details

Parameter	Value
Base model	Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled
Training GPU	NVIDIA A100 SXM4 80GB
Framework	Unsloth + TRL SFTTrainer
Dataset	OpenClaw tool-call examples (243 examples)
Training time	~102 minutes
Epochs	3
Steps	48
Final loss	0.7066
LoRA rank	r=32, alpha=64, rsLoRA=True
LoRA dropout	0.05
LoRA targets	q/k/v/o/gate/up/down proj
Trainable params	159,383,552 (0.58%)
Context window	4096 tokens
Batch size	1 (effective: 8 with grad accum)
Learning rate	2e-4 (cosine schedule, 5% warmup)
Quantization	4-bit NF4 during training
Optimizer	AdamW 8-bit

What It Does

This adapter teaches the model the OpenClaw tool-calling format — a structured XML-style invocation pattern used by the OpenClaw AI agent platform:

xml

<function_calls>
<invoke name="TOOL_NAME">
<parameter name="PARAM_NAME">value</parameter>
</invoke>
</function_calls>

Supported tools in training data: exec, read, write, edit, web_search, web_fetch, browser, memory_search, memory_get, message, cron, nodes, image, pdf, sessions_spawn, session_status

Usage with llama.cpp / Ollama

bash

# Ollama (Q4_K_M)
ollama run hf.co/peterjohannmedina/Medina-Qwen3.5-27B-OpenClaw:Q4_K_M

# llama.cpp direct
./llama-cli -m Medina-Qwen3.5-27B-OpenClaw-Q4_K_M.gguf \
  --ctx-size 4096 -p "You are an AI assistant with access to tools..."

Usage with Transformers (LoRA adapter)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "peterjohannmedina/Medina-Qwen3.5-27B-OpenClaw")
tokenizer = AutoTokenizer.from_pretrained("peterjohannmedina/Medina-Qwen3.5-27B-OpenClaw")

Companion Model

For a smaller version that runs on M3 MacBook / 16GB systems:

Medina-Qwen3-14B-OpenClaw (Q4_K_M: 8.4 GB, Q8_0: 14.6 GB)

License

Apache 2.0 — same as the base model.