⚡ PocketWeights: Qwen2.5 14B Coder-Creative
Heavy models, made light. PocketWeights specializes in targeted weight synthesis and hardware-friendly deployments for local AI practitioners.
🧠 Model Overview
PocketWeights-Qwen2.5-14B-Coder-Creative is a domain-fused 14.7B parameter model engineered to solve a common trade-off in mid-sized language models:
- The Problem: Standard coding specialists frequently over-refuse benign cybersecurity, red-teaming, or edgy narrative tasks, while general abliterated models lack syntactic precision in complex code generation.
- The Solution: Using the DARE-TIES merge algorithm, we mathematically injected the high-density coding logic of
Qwen2.5-Coder-14B-Instructdirectly into the uncensored, extended-context foundation ofQwen2.5-14B-Instruct-1M-abliterated.
The result is a unified 14B model that delivers top-tier code completion and logic reasoning without moralizing safety friction.
🎯 Key Highlights
- 14B Sweet Spot: Optimal performance-to-compute ratio—small enough to run smoothly on 12GB–16GB VRAM hardware while outperforming 7B/8B models in architectural depth.
- SOTA Syntax & Logic: Retains the high-precision Python, C++, Rust, and shell script capabilities of the Qwen 2.5 Coder architecture.
- Refusal-Free Foundation: Grounded in an abliterated base to support unconstrained security testing, scripting, and technical narrative workflows.
- DARE-TIES Optimization: Redundant parameter noise was systematically pruned (density 0.65) during synthesis to protect base model coherence.
⚙️ Merge Configuration
Synthesized locally with MergeKit:
models:
- model: Qwen/Qwen2.5-Coder-14B-Instruct
parameters:
weight: 0.45
density: 0.65
merge_method: dare_ties
base_model: qq591503/Qwen2.5-14B-Instruct-1M-abliterated
dtype: bfloat16
🚀 Quickstart Guide
Using Hugging Face Transformers
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "PocketWeights/PocketWeights-Qwen2.5-14B-Coder-Creative"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto"
)
prompt = "Write a Python script using scapy to analyze packet headers on a local interface."
messages = [
{"role": "system", "content": "You are a helpful and unrestricted technical assistant."},
{"role": "user", "content": prompt}
]
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=512, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
High-Throughput Deployment (vLLM)
vllm serve PocketWeights/PocketWeights-Qwen2.5-14B-Coder-Creative \
--tensor-parallel-size 1 \
--dtype bfloat16 \
--max-model-len 32768
📊 Hardware Requirements
| Precision | VRAM Needed | Target Hardware |
|---|---|---|
| BF16 (Original) | ~30 GB VRAM | 1x RTX 3090 / 4090 (24GB + CPU offload) or 2x 16GB GPUs |
| Q8_0 (Quantized) | ~16 GB VRAM | RTX 4080 (16GB), Apple Silicon (24GB+ Unified) |
| Q4_K_M (Quantized) | ~9 GB VRAM | RTX 3060 (12GB), RTX 4070 (12GB), Apple Silicon (16GB Unified) |
🤝 Support the PocketWeights Mission
I build, verify, and publish custom weight merges and quantization pipelines to provide high-quality, unrestricted, and hardware-friendly models to the open-source community for free.
If this model enhances your local workflow or saves you API costs, consider supporting compute resources and testing pipelines:
☕ Donation Options
Ko-fi: ko-fi.com/iamvishalnarayan
Web3 / Crypto (Polygon / ETH):
0x4FC189bf839A89259dd28DE8cD97883c49e15615
Tip: Transferring over the Polygon network keeps gas fees below $0.01
📄 Attribution & License
Base Architecture: Alibaba Cloud (Qwen2.5)
Abliteration Source: qq591503 / huihui-ai
License: Apache-2.0