CoolFace
Modelpublic

umitaksoylu/lsda-3b-turkish-dev-GGUF

sourceHugging Faceapache-2.0updated 5d agoView on Hugging Face
0likes149downloads
Model Card

LSDA 3B Turkish Dev - GGUF Versions

This repository contains GGUF quantized versions of umitaksoylu/lsda-3b-turkish-dev, compiled using llama.cpp.

LSDA-3B-Turkish-Dev is a high-performance 3B parameter LLM (based on Qwen2.5 Architecture) specifically trained for modern full-stack software development with a deep focus on C#, SQL, and React. LSDA Supports English & Turkish languages.

![Visit_Official_Website](https://lsda.umitaksoylu.com)

📦 Available Files & Specs

File NameQuant MethodFile SizeRecommended RAM/VRAMPrimary Use Case
lsda-3b-turkish-dev-Q4_K_M.ggufQ4_K_M~2.1 GB≥ 4 GBMobile devices (iOS/Android), low VRAM GPUs (e.g. GTX 1650), or CPU-only local environments. Minimal quality degradation with optimal speed.
lsda-3b-turkish-dev-F16.ggufFP16~6.0 GB≥ 8 GBUncompressed reference tests and evaluations on regular CPU's or Apple Silicon (Metal).

🛠️ Supported Inference Engines

You can run these GGUF weights across a wide variety of edge, mobile, and server engines:

  • —llama.cpp (Native cross-platform execution)
  • —Ollama (Local REST server & CLI)
  • —MLC-LLM (Cross-platform GPU acceleration)
  • —ExecuTorch (PyTorch native mobile engine)
  • —MediaPipe GenAI (On-device deployment)
  • —RKNN-LLM (Rockchip NPU acceleration)
  • —TensorRT-LLM (NVIDIA Edge/Server acceleration)

📱 Device Compatibility & Benchmarks ("The Developer in Your Pocket")

Thanks to GGUF and 4-bit quantization support, LSDA-3B can run locally across a wide range of hardware, including smartphones, tablets, edge AI boards, and low-power Mini PCs.

🍎 Apple Ecosystem (Metal / Apple Neural Engine)

Device SeriesMemory (RAM)Expected Experience / Performance
iPhone 16 / 16 Plus / Pro / Pro Max8 GBIdeal local on-device experience
iPhone 15 Pro / 15 Pro Max8 GBAccelerated by Apple A17 Pro ANE hardware
iPhone 15, 14 (Base/Pro), 13 Pro6 GBStable & smooth performance with 4-bit (Q4) models
iPhone 13 (Base) / 13 Mini4 GBLightweight local usage with 4-bit (Q4) or aggressive quants
iPad Pro & iPad Air (M1 / M2 / M4)8 GB – 16 GBDesktop-grade local inference speeds
iPad Mini / iPad (A15 / A16 / A17)4 GB – 6 GBPortable, lightweight local coding assistant

🤖 Android Devices (Min. 6 GB, Recommended 8 GB+ RAM)

Chipset ArchitectureExample Devices / HardwarePerformance
Qualcomm Snapdragon 8 Gen 1 / 2 / 3 / 8 EliteGalaxy S23/S24, Xiaomi 13/14 (Hexagon NPU & Adreno GPU)25 – 45 tok/s
Qualcomm Snapdragon 7+ Gen 2 / 7 Gen 3Upper mid-range Android smartphonesFluid local inference
Samsung Exynos 2400 / 2200 / 1480AMD RDNA-based Xclipse GPU & NPU acceleration~20 – 35 tok/s
Huawei Kirin 9000 / 9000s / 9010Huawei Mate 60, Pura 70 (DaVinci NPU)Full hardware support
MediaTek Dimensity 9200 / 9300 / 9400 & 8300Flagship/Sub-flagship devices (APU NPU architecture)High power efficiency

🤖 Embedded, Robotics & Edge AI

Platform / HardwareSpecs / AccelerationInference Speed
NVIDIA Jetson Orin Nano (4GB / 8GB)TensorRT-LLM & llama.cpp support~30+ tok/s
NVIDIA Jetson Orin NX / AGX OrinIndustrial automation & real-time roboticsHigh throughput
Raspberry Pi 5 (4GB / 8GB)CPU execution via llama.cpp / Ollama (Q4KM)~5 – 8 tok/s
Orange Pi 5 / 5+ (RK3588)6 TOPS NPU acceleration via RKNN-LLM~12 – 18 tok/s
Intel N100 / N97 / Core i3-N3056W–15W Ultra Low-Power x86 Mini PCs / IoT Gateways~10 – 15 tok/s

🚀 Quick Start

1. Running via llama.cpp CLI

bash
# Download the quantized GGUF file
huggingface-cli download umitaksoylu/lsda-3b-turkish-dev-GGUF lsda-3b-turkish-dev-Q4_K_M.gguf --local-dir .

# Run with Qwen 2.5 chat template
./llama-cli -m lsda-3b-turkish-dev-Q4_K_M.gguf \
  --chat-template qwen2 \
  -p "Write a C# DTO class and a corresponding React interface for a User Profile." \
  -n 512