Qwable-EX

A personal experimental linear merge of strong Qwen3.5-9B lineage models focused on structured reasoning, agentic coding, and tool-use behavior.

Qwable-EX combines the deliberate, trace-style reasoning of Empero’s Qwable fine-tune with complementary strengths from Qworus-V2 and Qwepus via a simple linear merge (Mergekit). It is intended as a research / experimentation checkpoint rather than a production model.

Model Details

Property	Value
Base architecture	Qwen3.5-9B (dense, hybrid Gated DeltaNet + full attention)
Parameter count	~9B
Merge method	Linear (Mergekit)
Primary lineage	Qwen3.5-9B → Qwable-style reasoning + complementary merges
Languages	Primarily English
License	MIT (this merge) — parent models are mostly Apache-2.0; see provenance notes

Component Models

• empero-ai/Qwable-9B-Claude-Fable-5
  Full-parameter SFT of Qwen3.5-9B on Claude Fable 5 reasoning/coding traces + a small GPT-5.5 terminal/agent set. Strong structured <think> reasoning and agentic coding style. Vision tower was frozen (text-only fine-tune).

• DarkKitsune/Qwen3.5-9B-Qworus-V2
  50/50 DARE-TIES merge of empero-ai/Qwen3.8-9B-Distill and ornith-ai/Ornith-1.5-9B. Hybrid reasoning model aimed at coding, tool calls, planning, and general question answering. Not uncensored by default.

• Netuoso/qwepus
  Additional Qwen-based component included in the linear merge.

All components ultimately descend from the Qwen3.5-9B family (natively multimodal architecture with long context support).

Intended Uses

• Experimental local inference for structured reasoning and coding tasks
• Exploring how linear merges of Qwable-style + hybrid reasoning models behave
• Agentic / tool-use style prompting (results will vary)
• Research and personal experimentation

Not intended for production deployment, high-stakes decision making, or use cases requiring strong safety guarantees without additional filtering.

How to Use

This is a standard Transformers-compatible checkpoint. Example (text-only):

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "YOUR_USERNAME/Qwable-EX"  # replace with actual repo

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Write a clean Python function that merges two sorted lists."}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
    repetition_penalty=1.05,
)

print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))