This repo exists to test 100 open-source models, document the results, and keep a set of ready, verified models to reuse in future projects.
Test environment: laptop, NVIDIA GTX 1650 (4GB VRAM) — so the tests mostly cover small/quantized models. This shows which models are practically usable under limited GPU resources (typical of many small-to-mid-size projects in the Uzbekistan market).
Categories: LLM, CV (Computer Vision), TTS (Text-to-Speech), STT (Speech-to-Text) — with special attention to Uzbek language support.
| # | Model | Category | Which project need it fits | GPU required | Estimated monthly cost | Same server as backend? | Details |
|---|---|---|---|---|---|---|---|
| 1 | YOLOv8n | CV | Camera-based monitoring, presence detection, object counting | No | $0 (runs on existing server) | ✅ Yes | README |
| 2 | YOLO11n | CV | Smart Desk Focus & Distraction Monitor (Workplace ergonomics) | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 3 | YOLO11s | CV | Retail & Checkout Queue Length Monitor (Retail flow analytics) | No | $0 (CPU / low-cost VPS) | ✅ Yes | README |
| 4 | YOLO11m | CV | Industrial Hazard & Heavy Machinery Guardian (Workplace safety) | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 5 | YOLOv10n | CV | Perimeter Intrusion & Tripwire Guardian (NMS-free security) | No | $0 (runs on existing server) | ✅ Yes | README |
| 6 | YOLOv10s | CV | Smart Parking Bay & Loitering Patrol (Smart cities & surveillance) | No | $0 (CPU / low-cost VPS) | ✅ Yes | README |
| 7 | YOLOv9t | CV | Highway Traffic Flow Counter (PGI dual-head speed & volume) | No | $0 (CPU / edge SBC) | ✅ Yes | README |
| 8 | YOLOv9s | CV | Pedestrian Crosswalk & Jaywalking Guardian (Vision Zero alerts) | No | $0 (CPU / edge node) | ✅ Yes | README |
| 9 | YOLOv8s | CV | Retail Loss Prevention Baggage Tracker (Anti-theft shopper audit) | No | $0 (CPU / edge server) | ✅ Yes | README |
| 10 | YOLOv8m | CV | Crowd Spatial Density & Cluster Analyzer (Public event safety) | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 11 | YOLOv7-tiny | CV | Commercial Facility Foot-Traffic Counter (Doorway flow monitor) | No | $0 (runs on existing server) | ✅ Yes | README |
| 12 | YOLOv6n | CV | Warehouse Conveyor Belt Throughput Monitor (Industrial sorting) | No | $0 (CPU / embedded device) | ✅ Yes | README |
| 13 | YOLOv6s | CV | Freight Terminal Dock Bay Occupancy Inspector (Intermodal audit) | No | $0 (CPU / low-cost VPS) | ✅ Yes | README |
| 14 | YOLOv5n | CV | Smart Office Energy Occupancy Guardian (HVAC / lighting savings) | No | $0 (CPU / edge IoT) | ✅ Yes | README |
| 15 | YOLOv5s | CV | Urban Intersection Safety Analyzer (Vehicle/pedestrian risk) | No | $0 (CPU / edge box) | ✅ Yes | README |
| 16 | YOLOv5m | CV | Commercial Fleet & Logistics Yard Dispatcher (Heavy transport) | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 17 | YOLOv4-tiny | CV | Edge Micro-UAV Aerial Target Recon (Low-power drone SAR) | No | $0 (pure OpenCV DNN CPU) | ✅ Yes | README |
| 18 | YOLOv3-tiny | CV | Legacy CPU Vehicle Parking Gate Actuator (Boom barrier trigger) | No | $0 (runs on legacy x86/Atom) | ✅ Yes | README |
| 19 | Depth-Anything-V2 | CV | Monocular 3D Depth Estimation, LiDAR-free spatial distance | Optional | $0 (CPU / edge-ready) | ✅ Yes | README |
| 20 | YOLO11n-Pose | CV | 17-Keypoint Human Skeleton & Workplace Posture / Fitness Tracker | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 21 | YOLO11n-Seg | CV | Multi-Class Instance Segmentation & Pixel-Accurate Object Masking | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 22 | YOLO-World (v2) | CV | Open-Vocabulary Zero-Shot Detection (Arbitrary prompt querying) | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 23 | OpenCV-Stream | CV | Zero-latency RTSP/Webcam stream ingestion & frame telemetry | No | $0 (pure CPU pipeline) | ✅ Yes | README |
| 24 | YOLO-Office-Detection | CV | Workplace presence, IT equipment & distraction alert detector | No | $0 (runs on edge/CPU) | ✅ Yes | README |
| 25 | InsightFace (ArcFace) | CV | SOTA SCRFD + ArcFace ResNet-50 biometric face identification | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 26 | YOLO-Pose-Sleeping | CV | 17-Keypoint skeleton workplace fatigue, posture & sleep analyzer | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 27 | ByteTrack-MOT | CV | Multi-Object Tracking (MOT), persistent ID assignment & trajectories | No | $0 (CPU Kalman/Hungarian) | ✅ Yes | README |
| 28 | BGE-M3 | LLM | World SOTA multilingual semantic embedding (1024-d, 8192 context, Uzbek+100 lang) | No | $0 (runs on CPU, ~20ms latency) | ✅ Yes | README |
| 29 | BGE-Reranker-v2-M3 | LLM | Cross-Encoder RAG reranker (Full Cross-Attention, false-positive elimination) | No | $0 (runs on CPU, ~30ms latency) | ✅ Yes | README |
| 30 | DeepSeek-Coder-V2-Lite | LLM | Mixture-of-Experts (MoE) 16B/2.4B active code model, 338 languages, 128k context | Optional | $0–$15 (runs on CPU / 8GB GPU) | ✅ Yes | README |
| 31 | DeepSeek-R1-1.5B | LLM | Offline reasoning engine with Chain-of-Thought () for math/logic | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 32 | DeepSeek-R1-Llama-8B | LLM | Deep reasoning, proof by contradiction & logic deduction (Meta Llama-3.1) | Recommended | $0–$25 (runs on CPU / 8GB GPU) | ✅ Yes | README |
| 33 | Gemma-2-2B | LLM | Google flagship 2.6B lightweight model, Sliding Window Attention, soft-capping | No | $0 (runs on CPU / 2GB RAM) | ✅ Yes | README |
| 34 | Granite-3.0-2B | LLM | IBM enterprise foundation model, 12T tokens, 100% Apache 2.0 unencumbered license | No | $0 (runs on CPU / 2GB RAM) | ✅ Yes | README |
| 35 | Hermes-3-Llama-3.2-3B | LLM | Autonomous agent execution, function calling <tools>, structured JSON schema | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 36 | Llama-3.1-8B | LLM | Meta industry-standard 8B foundation, 15T tokens, 128k context, enterprise RAG | Recommended | $0–$25 (runs on CPU / 8GB GPU) | ✅ Yes | README |
| 37 | Llama-3.2-1B | LLM | Meta official smartphone/IoT edge engine, 128k context, ~800MB RAM | No | $0 (runs on CPU / 1GB RAM) | ✅ Yes | README |
| 38 | Llama-Guard-3-1B | LLM | Meta AI safety & content moderation classifier (S1-S14 hazard policies) | No | $0 (runs on CPU, ~50ms audit) | ✅ Yes | README |
| 39 | Mistral-7B-v0.3 | LLM | European SOTA 7B flagship, Tekken tokenizer, 32k context, native function calling | Recommended | $0–$25 (runs on CPU / 6GB GPU) | ✅ Yes | README |
| 40 | Moondream2 | LLM | Ultra-lightweight Tiny VLM (~1.86B), instant edge visual Q&A & image captioning | No | $0 (runs on CPU / 2GB RAM) | ✅ Yes | README |
| 41 | NLLB-200-Distilled | LLM | Meta 200-language direct neural machine translation (Uzbek Latin & Cyrillic) | No | $0 (runs on CPU, ~200ms latency) | ✅ Yes | README |
| 42 | Phi-3.5-mini | LLM | Microsoft 3.8B high-density synthetic reasoning, 128k context, math/logic SOTA | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 43 | Qwen2-VL-2B | LLM | Vision-Language multimodal understanding (VLM), image OCR, chart & document VQA | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 44 | Qwen2.5-1.5B | LLM | Local generative assistant, FAQ bot, structured data extractor | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 45 | Qwen2.5-3B | LLM | Alibaba 3B sweet-spot powerhouse, 18T tokens, deep Uzbek & Turkic multilingual | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 46 | Qwen2.5-Coder-1.5B | LLM | Local Copilot, code generation, refactoring & syntax repair (5.5T tokens) | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 47 | SmolLM2-1.7B | LLM | HuggingFace 11T-token curated synthetic model, ultra-compact mobile/edge | No | $0 (runs on CPU / 2GB RAM) | ✅ Yes | README |
| 48 | StarCoder2-3B | LLM | BigCode enterprise-safe code generation (The Stack v2, 100% Permissive) | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 49 | Bark-Small | TTS | Expressive generative audio, audiobooks, dialogue with laughter and sighs | Optional | $0–$15 (CPU / entry GPU) | ✅ Yes | README |
| 50 | ChatTTS | TTS | Conversational dialogue synthesis optimized for LLM agents with laughter and filler words | No | $0 (runs on CPU) | ✅ Yes | README |
| 51 | Coqui-TTS-VITS | TTS | End-to-end conditional variational autoencoder for low-latency voice bots | No | $0 (CPU-ready) | ✅ Yes | README |
| 52 | Coqui-XTTS-v2 | TTS | Zero-shot voice cloning across 17+ languages from 3-second reference audio | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 53 | CosyVoice-300M | TTS | Multilingual zero-shot voice cloning and emotional conversational synthesis | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 54 | E2-TTS | TTS | Duration-free non-autoregressive speech synthesis with instant alignment | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 55 | eSpeak-NG | TTS | Microcontroller / embedded formant synthesizer with microsecond latency (<10MB RAM) | No | $0 (runs on $1 micro-server) | ✅ Yes | README |
| 56 | F5-TTS | TTS | Non-autoregressive Flow Matching speech synthesis with rapid inference | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 57 | FastSpeech2 | TTS | Deterministic ultra-fast non-autoregressive speech synthesis without word skipping | No | $0 (runs on CPU) | ✅ Yes | README |
| 58 | Fish-Speech-1.5 | TTS | Dual-autoregressive multi-lingual voice generator with low memory consumption | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 59 | Glow-TTS | TTS | Generative flow parallel acoustic model with monotonic alignment search | No | $0 (CPU-ready) | ✅ Yes | README |
| 60 | Matcha-TTS | TTS | Fast, high-quality, lightweight non-autoregressive ODE flow matching | No | $0 (runs smoothly on CPU) | ✅ Yes | README |
| 61 | MetaVoice-1B | TTS | Conversational 1.2B foundation model trained on 100k hours of expressive dialogue | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 62 | MMS-TTS-KAZ | TTS | Sister Turkic language TTS for Central Asian regional localization and testing | No | $0 (CPU-ready) | ✅ Yes | README |
| 63 | MMS-TTS-UZB (Latin) | TTS | Official Uzbek Latin text-to-speech audio synthesis (Government & Fintech) | No | $0 (runs on existing server / CPU) | ✅ Yes | README |
| 64 | MMS-TTS-UZB | TTS | Uzbek language text-to-speech audio synthesizer (Voice bots) | No | $0 (VITS architecture on CPU) | ✅ Yes | README |
| 65 | OpenVoice-v2 | TTS | Instant versatile voice cloning with independent control over tone color and emotion | No | $0 (CPU-ready) | ✅ Yes | README |
| 66 | Parler-TTS-Mini | TTS | Controllable speech generation guided by natural language prompts (gender, tone, pace) | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 67 | Piper-TTS | TTS | Ultra-fast local neural TTS for embedded systems, Raspberry Pi, and low-cost CPU VPS | No | $0 (runs on edge / $3 VPS) | ✅ Yes | README |
| 68 | SeamlessM4T-TTS | TTS | Expressive multilingual multi-task text-to-speech across 35+ languages | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 69 | Sherpa-ONNX-TTS | TTS | Next-gen Kaldi embedded offline neural TTS for Android, iOS, and Linux edge SBCs | No | $0 (runs on edge / CPU) | ✅ Yes | README |
| 70 | SpeechT5-TTS | TTS | Unified encoder-decoder framework with customizable x-vector speaker embeddings | No | $0 (CPU-ready) | ✅ Yes | README |
| 71 | StyleTTS2 | TTS | Human-level speech synthesis using style diffusion and adversarial training | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 72 | Tacotron2 | TTS | Industry-classic recurrent sequence-to-sequence Mel-spectrogram generator | No | $0 (CPU-ready) | ✅ Yes | README |
| 73 | Tortoise-TTS | TTS | Studio-quality multi-voice narration, audiobooks, and luxury voice generation | Recommended | $15–$30 (GPU required for speed) | ⚠️ Dedicated recommended | README |
| 74 | VALL-E-X | TTS | Zero-shot cross-lingual speech synthesis and speech-to-speech translation | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 75 | Conformer-CTC | STT | Enterprise call-center ASR combining self-attention with depthwise convolutions | No | $0 (runs on CPU) | ✅ Yes | README |
| 76 | Data2Vec-Audio-Large | STT | Unified multi-modal self-supervised architecture applied to speech recognition | No | $0 (runs on CPU) | ✅ Yes | README |
| 77 | FasterWhisper | STT | Fast audio transcription & voice command parsing (Uzbek & Multi) | Optional | $0–$10 (CTranslate2 on CPU) | ✅ Yes | README |
| 78 | FunASR-Paraformer | STT | Non-autoregressive industrial speech recognition with continuous streaming support | No | $0 (runs on CPU) | ✅ Yes | README |
| 79 | HuBERT-Large | STT | Self-supervised speech representation model with k-means acoustic clustering | No | $0 (runs on CPU) | ✅ Yes | README |
| 80 | Insanely-Fast-Whisper | STT | Batched ultra-fast pipeline inference using Flash Attention 2 and Hugging Face Optimum | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 81 | MMS-1B-All | STT | Meta 1,400-language foundation ASR model for universal speech recognition | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 82 | Moonshine-Base | STT | Low-latency live microphone speech transcription for real-time captions | No | $0 (runs on CPU) | ✅ Yes | README |
| 83 | Moonshine-Tiny | STT | Instant edge streaming speech recognition without 30-second chunk latency | No | $0 (runs on microcontrollers) | ✅ Yes | README |
| 84 | NeMo-Canary-1B | STT | Multitask speech transcription and translation with punctuation and capitalization | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 85 | SeamlessM4T-STT | STT | Multilingual automatic speech recognition and speech-to-text translation across 100+ languages | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 86 | SenseVoice-Small | STT | Ultra-fast speech recognition (<100ms latency) with rich audio event and emotion detection | No | $0 (runs on CPU) | ✅ Yes | README |
| 87 | Sherpa-ONNX-STT | STT | Offline edge embedded speech recognition for Linux, Android, and IoT SBCs | No | $0 (runs on edge / $3 VPS) | ✅ Yes | README |
| 88 | Silero-STT | STT | Enterprise-grade compact models running on single CPU thread (<30MB) with real-time streaming | No | $0 (runs on micro-server) | ✅ Yes | README |
| 89 | Vosk-API-UZ | STT | Lightweight Kaldi-based offline speech recognizer (~50MB model) for mobile and Raspberry Pi | No | $0 (runs on CPU / embedded) | ✅ Yes | README |
| 90 | Wav2Vec2-XLSR-UZ | STT | End-to-end self-supervised acoustic CTC model fine-tuned for Uzbek speech | No | $0 (runs on CPU) | ✅ Yes | README |
| 91 | WavLM-Large | STT | Speech recognition with native background denoising and speaker verification | No | $0 (runs on CPU) | ✅ Yes | README |
| 92 | Whisper-Base | STT | Optimal balance of edge speed and acceptable transcription accuracy for bots | No | $0 (runs on existing server) | ✅ Yes | README |
| 93 | Whisper-Diarization | STT | Multi-speaker meeting transcription with speaker identification ('Who Spoke When') | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 94 | Whisper-Large-v3-Turbo | STT | High-accuracy enterprise transcription at 8x speed (pruned 4-layer decoder) | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 95 | Whisper-Medium | STT | Production-grade call-center analytics and complex multi-speaker transcription | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 96 | Whisper-Small-UZ | STT | Domain-fine-tuned Uzbek speech recognition on Common Voice Uzbek corpus | No | $0 (runs on CPU) | ✅ Yes | README |
| 97 | Whisper-Timestamped | STT | Accurate word-level timestamping and subtitle synchronization without hallucinations | No | $0 (runs on CPU) | ✅ Yes | README |
| 98 | Whisper-Tiny | STT | Ultra-lightweight real-time transcription on low-end CPUs and IoT hardware | No | $0 (runs on $4 VPS) | ✅ Yes | README |
| 99 | Whisper.cpp | STT | High-performance pure C/C++ ASR inference with zero Python runtime overhead | No | $0 (runs on bare-metal CPU) | ✅ Yes | README |
| 100 | Zipformer-Transducer | STT | Next-gen Kaldi Zipformer multi-rate transducer with exceptional parameter efficiency | No | $0 (runs on CPU) | ✅ Yes | README |
(100 of 100 models reviewed, scaffolded and verified — 100% COMPLETE)
100-opensource-models-review/
├── README.md # this file — overall index
├── _template/ # template files and rules for adding a new model
├── cv/
│ ├── requirements.txt # unified master dependencies for all 27 CV models
│ ├── bytetrack-mot/
│ ├── depth-anything-v2-small/
│ ├── insightface-arcface/
│ ├── opencv-video-stream/
│ ├── yolo-object-detection/
│ ├── yolo-pose-sleeping/
│ ├── yolo11m/
│ ├── yolo11n/
│ ├── yolo11n-pose/
│ ├── yolo11n-seg/
│ ├── yolo11s/
│ ├── yolov10n/
│ ├── yolov10s/
│ ├── yolov3-tiny/
│ ├── yolov4-tiny/
│ ├── yolov5m/
│ ├── yolov5n/
│ ├── yolov5s/
│ ├── yolov6n/
│ ├── yolov6s/
│ ├── yolov7-tiny/
│ ├── yolov8m/
│ ├── yolov8n/
│ ├── yolov8s/
│ ├── yolov8s-worldv2/
│ ├── yolov9s/
│ ├── yolov9t/
├── llm/
│ ├── DeepSeek-Coder-V2-Lite-Instruct-GGUF/
│ ├── DeepSeek-R1-Distill-Llama-8B-GGUF/
│ ├── DeepSeek-R1-Distill-Qwen-1.5B-GGUF/
│ ├── Gemma-2-2B-Instruct-GGUF/
│ ├── Granite-3.0-2B-Instruct-GGUF/
│ ├── Hermes-3-Llama-3.2-3B-GGUF/
│ ├── Llama-3.1-8B-Instruct-GGUF/
│ ├── Llama-3.2-1B-Instruct-GGUF/
│ ├── Llama-Guard-3-1B-GGUF/
│ ├── Mistral-7B-Instruct-v0.3-GGUF/
│ ├── Moondream2-GGUF/
│ ├── NLLB-200-Distilled-600M/
│ ├── Phi-3.5-mini-instruct-GGUF/
│ ├── Qwen2-VL-2B-Instruct-GGUF/
│ ├── Qwen2.5-1.5B-Instruct-GGUF/
│ ├── Qwen2.5-3B-Instruct-GGUF/
│ ├── Qwen2.5-Coder-1.5B-Instruct-GGUF/
│ ├── SmolLM2-1.7B-Instruct-GGUF/
│ ├── StarCoder2-3B-GGUF/
│ ├── bge-m3/
│ ├── bge-reranker-v2-m3/
├── tts/
│ ├── bark-small/
│ ├── chattts/
│ ├── coqui-tts-vits/
│ ├── coqui-xtts-v2/
│ ├── cosyvoice-300m/
│ ├── e2-tts/
│ ├── espeak-ng/
│ ├── f5-tts/
│ ├── fastspeech2/
│ ├── fish-speech-1.5/
│ ├── glow-tts/
│ ├── matcha-tts/
│ ├── meta-voice-1b/
│ ├── mms-tts-kaz/
│ ├── mms-tts-uzb/
│ ├── mms-tts-uzb-latin/
│ ├── openvoice-v2/
│ ├── parler-tts-mini/
│ ├── piper-tts/
│ ├── seamless-m4t-tts/
│ ├── sherpa-onnx-offline-tts/
│ ├── speecht5-tts/
│ ├── styletts2/
│ ├── tacotron2/
│ ├── tortoise-tts/
│ ├── valle-x/
├── stt/
│ ├── FasterWhisper/
│ ├── conformer-ctc/
│ ├── data2vec-audio-large/
│ ├── funasr-paraformer/
│ ├── hubert-large-ls960/
│ ├── insanely-fast-whisper/
│ ├── mms-1b-all/
│ ├── moonshine-base/
│ ├── moonshine-tiny/
│ ├── nemo-canary-1b/
│ ├── seamless-m4t-stt/
│ ├── sensevoice-small/
│ ├── sherpa-onnx-offline-stt/
│ ├── silero-stt/
│ ├── vosk-api-uz/
│ ├── wav2vec2-large-xlsr-uz/
│ ├── wavlm-large/
│ ├── whisper-base/
│ ├── whisper-cpp/
│ ├── whisper-diarization/
│ ├── whisper-large-v3-turbo/
│ ├── whisper-medium/
│ ├── whisper-small-uz/
│ ├── whisper-timestamped/
│ ├── whisper-tiny/
│ ├── zipformer-transducer/
└── benchmark_scripts/ # shared benchmark scripts per category
Each model folder is self-contained — it has its own Dockerfile, run_benchmarks.py, and README.md.
cv/venv-cv virtual environment (pip install -r cv/requirements.txt) with full local webcam (--source 0), GUI preview, or --headless batch processing.docker compose up <model-name> --build) or locally via Python.Rules and template files for adding a new model — _template/HOW_TO_ADD_A_MODEL.md.
This repo exists to test 100 open-source models, document the results, and keep a set of ready, verified models to reuse in future projects.
Test environment: laptop, NVIDIA GTX 1650 (4GB VRAM) — so the tests mostly cover small/quantized models. This shows which models are practically usable under limited GPU resources (typical of many small-to-mid-size projects in the Uzbekistan market).
Categories: LLM, CV (Computer Vision), TTS (Text-to-Speech), STT (Speech-to-Text) — with special attention to Uzbek language support.
| # | Model | Category | Which project need it fits | GPU required | Estimated monthly cost | Same server as backend? | Details |
|---|---|---|---|---|---|---|---|
| 1 | YOLOv8n | CV | Camera-based monitoring, presence detection, object counting | No | $0 (runs on existing server) | ✅ Yes | README |
| 2 | YOLO11n | CV | Smart Desk Focus & Distraction Monitor (Workplace ergonomics) | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 3 | YOLO11s | CV | Retail & Checkout Queue Length Monitor (Retail flow analytics) | No | $0 (CPU / low-cost VPS) | ✅ Yes | README |
| 4 | YOLO11m | CV | Industrial Hazard & Heavy Machinery Guardian (Workplace safety) | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 5 | YOLOv10n | CV | Perimeter Intrusion & Tripwire Guardian (NMS-free security) | No | $0 (runs on existing server) | ✅ Yes | README |
| 6 | YOLOv10s | CV | Smart Parking Bay & Loitering Patrol (Smart cities & surveillance) | No | $0 (CPU / low-cost VPS) | ✅ Yes | README |
| 7 | YOLOv9t | CV | Highway Traffic Flow Counter (PGI dual-head speed & volume) | No | $0 (CPU / edge SBC) | ✅ Yes | README |
| 8 | YOLOv9s | CV | Pedestrian Crosswalk & Jaywalking Guardian (Vision Zero alerts) | No | $0 (CPU / edge node) | ✅ Yes | README |
| 9 | YOLOv8s | CV | Retail Loss Prevention Baggage Tracker (Anti-theft shopper audit) | No | $0 (CPU / edge server) | ✅ Yes | README |
| 10 | YOLOv8m | CV | Crowd Spatial Density & Cluster Analyzer (Public event safety) | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 11 | YOLOv7-tiny | CV | Commercial Facility Foot-Traffic Counter (Doorway flow monitor) | No | $0 (runs on existing server) | ✅ Yes | README |
| 12 | YOLOv6n | CV | Warehouse Conveyor Belt Throughput Monitor (Industrial sorting) | No | $0 (CPU / embedded device) | ✅ Yes | README |
| 13 | YOLOv6s | CV | Freight Terminal Dock Bay Occupancy Inspector (Intermodal audit) | No | $0 (CPU / low-cost VPS) | ✅ Yes | README |
| 14 | YOLOv5n | CV | Smart Office Energy Occupancy Guardian (HVAC / lighting savings) | No | $0 (CPU / edge IoT) | ✅ Yes | README |
| 15 | YOLOv5s | CV | Urban Intersection Safety Analyzer (Vehicle/pedestrian risk) | No | $0 (CPU / edge box) | ✅ Yes | README |
| 16 | YOLOv5m | CV | Commercial Fleet & Logistics Yard Dispatcher (Heavy transport) | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 17 | YOLOv4-tiny | CV | Edge Micro-UAV Aerial Target Recon (Low-power drone SAR) | No | $0 (pure OpenCV DNN CPU) | ✅ Yes | README |
| 18 | YOLOv3-tiny | CV | Legacy CPU Vehicle Parking Gate Actuator (Boom barrier trigger) | No | $0 (runs on legacy x86/Atom) | ✅ Yes | README |
| 19 | Depth-Anything-V2 | CV | Monocular 3D Depth Estimation, LiDAR-free spatial distance | Optional | $0 (CPU / edge-ready) | ✅ Yes | README |
| 20 | YOLO11n-Pose | CV | 17-Keypoint Human Skeleton & Workplace Posture / Fitness Tracker | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 21 | YOLO11n-Seg | CV | Multi-Class Instance Segmentation & Pixel-Accurate Object Masking | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 22 | YOLO-World (v2) | CV | Open-Vocabulary Zero-Shot Detection (Arbitrary prompt querying) | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 23 | OpenCV-Stream | CV | Zero-latency RTSP/Webcam stream ingestion & frame telemetry | No | $0 (pure CPU pipeline) | ✅ Yes | README |
| 24 | YOLO-Office-Detection | CV | Workplace presence, IT equipment & distraction alert detector | No | $0 (runs on edge/CPU) | ✅ Yes | README |
| 25 | InsightFace (ArcFace) | CV | SOTA SCRFD + ArcFace ResNet-50 biometric face identification | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 26 | YOLO-Pose-Sleeping | CV | 17-Keypoint skeleton workplace fatigue, posture & sleep analyzer | No | $0 (CPU / edge-ready) | ✅ Yes | README |
| 27 | ByteTrack-MOT | CV | Multi-Object Tracking (MOT), persistent ID assignment & trajectories | No | $0 (CPU Kalman/Hungarian) | ✅ Yes | README |
| 28 | BGE-M3 | LLM | World SOTA multilingual semantic embedding (1024-d, 8192 context, Uzbek+100 lang) | No | $0 (runs on CPU, ~20ms latency) | ✅ Yes | README |
| 29 | BGE-Reranker-v2-M3 | LLM | Cross-Encoder RAG reranker (Full Cross-Attention, false-positive elimination) | No | $0 (runs on CPU, ~30ms latency) | ✅ Yes | README |
| 30 | DeepSeek-Coder-V2-Lite | LLM | Mixture-of-Experts (MoE) 16B/2.4B active code model, 338 languages, 128k context | Optional | $0–$15 (runs on CPU / 8GB GPU) | ✅ Yes | README |
| 31 | DeepSeek-R1-1.5B | LLM | Offline reasoning engine with Chain-of-Thought () for math/logic | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 32 | DeepSeek-R1-Llama-8B | LLM | Deep reasoning, proof by contradiction & logic deduction (Meta Llama-3.1) | Recommended | $0–$25 (runs on CPU / 8GB GPU) | ✅ Yes | README |
| 33 | Gemma-2-2B | LLM | Google flagship 2.6B lightweight model, Sliding Window Attention, soft-capping | No | $0 (runs on CPU / 2GB RAM) | ✅ Yes | README |
| 34 | Granite-3.0-2B | LLM | IBM enterprise foundation model, 12T tokens, 100% Apache 2.0 unencumbered license | No | $0 (runs on CPU / 2GB RAM) | ✅ Yes | README |
| 35 | Hermes-3-Llama-3.2-3B | LLM | Autonomous agent execution, function calling <tools>, structured JSON schema | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 36 | Llama-3.1-8B | LLM | Meta industry-standard 8B foundation, 15T tokens, 128k context, enterprise RAG | Recommended | $0–$25 (runs on CPU / 8GB GPU) | ✅ Yes | README |
| 37 | Llama-3.2-1B | LLM | Meta official smartphone/IoT edge engine, 128k context, ~800MB RAM | No | $0 (runs on CPU / 1GB RAM) | ✅ Yes | README |
| 38 | Llama-Guard-3-1B | LLM | Meta AI safety & content moderation classifier (S1-S14 hazard policies) | No | $0 (runs on CPU, ~50ms audit) | ✅ Yes | README |
| 39 | Mistral-7B-v0.3 | LLM | European SOTA 7B flagship, Tekken tokenizer, 32k context, native function calling | Recommended | $0–$25 (runs on CPU / 6GB GPU) | ✅ Yes | README |
| 40 | Moondream2 | LLM | Ultra-lightweight Tiny VLM (~1.86B), instant edge visual Q&A & image captioning | No | $0 (runs on CPU / 2GB RAM) | ✅ Yes | README |
| 41 | NLLB-200-Distilled | LLM | Meta 200-language direct neural machine translation (Uzbek Latin & Cyrillic) | No | $0 (runs on CPU, ~200ms latency) | ✅ Yes | README |
| 42 | Phi-3.5-mini | LLM | Microsoft 3.8B high-density synthetic reasoning, 128k context, math/logic SOTA | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 43 | Qwen2-VL-2B | LLM | Vision-Language multimodal understanding (VLM), image OCR, chart & document VQA | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 44 | Qwen2.5-1.5B | LLM | Local generative assistant, FAQ bot, structured data extractor | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 45 | Qwen2.5-3B | LLM | Alibaba 3B sweet-spot powerhouse, 18T tokens, deep Uzbek & Turkic multilingual | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 46 | Qwen2.5-Coder-1.5B | LLM | Local Copilot, code generation, refactoring & syntax repair (5.5T tokens) | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 47 | SmolLM2-1.7B | LLM | HuggingFace 11T-token curated synthetic model, ultra-compact mobile/edge | No | $0 (runs on CPU / 2GB RAM) | ✅ Yes | README |
| 48 | StarCoder2-3B | LLM | BigCode enterprise-safe code generation (The Stack v2, 100% Permissive) | No | $0 (runs on CPU / 4GB GPU) | ✅ Yes | README |
| 49 | Bark-Small | TTS | Expressive generative audio, audiobooks, dialogue with laughter and sighs | Optional | $0–$15 (CPU / entry GPU) | ✅ Yes | README |
| 50 | ChatTTS | TTS | Conversational dialogue synthesis optimized for LLM agents with laughter and filler words | No | $0 (runs on CPU) | ✅ Yes | README |
| 51 | Coqui-TTS-VITS | TTS | End-to-end conditional variational autoencoder for low-latency voice bots | No | $0 (CPU-ready) | ✅ Yes | README |
| 52 | Coqui-XTTS-v2 | TTS | Zero-shot voice cloning across 17+ languages from 3-second reference audio | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 53 | CosyVoice-300M | TTS | Multilingual zero-shot voice cloning and emotional conversational synthesis | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 54 | E2-TTS | TTS | Duration-free non-autoregressive speech synthesis with instant alignment | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 55 | eSpeak-NG | TTS | Microcontroller / embedded formant synthesizer with microsecond latency (<10MB RAM) | No | $0 (runs on $1 micro-server) | ✅ Yes | README |
| 56 | F5-TTS | TTS | Non-autoregressive Flow Matching speech synthesis with rapid inference | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 57 | FastSpeech2 | TTS | Deterministic ultra-fast non-autoregressive speech synthesis without word skipping | No | $0 (runs on CPU) | ✅ Yes | README |
| 58 | Fish-Speech-1.5 | TTS | Dual-autoregressive multi-lingual voice generator with low memory consumption | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 59 | Glow-TTS | TTS | Generative flow parallel acoustic model with monotonic alignment search | No | $0 (CPU-ready) | ✅ Yes | README |
| 60 | Matcha-TTS | TTS | Fast, high-quality, lightweight non-autoregressive ODE flow matching | No | $0 (runs smoothly on CPU) | ✅ Yes | README |
| 61 | MetaVoice-1B | TTS | Conversational 1.2B foundation model trained on 100k hours of expressive dialogue | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 62 | MMS-TTS-KAZ | TTS | Sister Turkic language TTS for Central Asian regional localization and testing | No | $0 (CPU-ready) | ✅ Yes | README |
| 63 | MMS-TTS-UZB (Latin) | TTS | Official Uzbek Latin text-to-speech audio synthesis (Government & Fintech) | No | $0 (runs on existing server / CPU) | ✅ Yes | README |
| 64 | MMS-TTS-UZB | TTS | Uzbek language text-to-speech audio synthesizer (Voice bots) | No | $0 (VITS architecture on CPU) | ✅ Yes | README |
| 65 | OpenVoice-v2 | TTS | Instant versatile voice cloning with independent control over tone color and emotion | No | $0 (CPU-ready) | ✅ Yes | README |
| 66 | Parler-TTS-Mini | TTS | Controllable speech generation guided by natural language prompts (gender, tone, pace) | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 67 | Piper-TTS | TTS | Ultra-fast local neural TTS for embedded systems, Raspberry Pi, and low-cost CPU VPS | No | $0 (runs on edge / $3 VPS) | ✅ Yes | README |
| 68 | SeamlessM4T-TTS | TTS | Expressive multilingual multi-task text-to-speech across 35+ languages | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 69 | Sherpa-ONNX-TTS | TTS | Next-gen Kaldi embedded offline neural TTS for Android, iOS, and Linux edge SBCs | No | $0 (runs on edge / CPU) | ✅ Yes | README |
| 70 | SpeechT5-TTS | TTS | Unified encoder-decoder framework with customizable x-vector speaker embeddings | No | $0 (CPU-ready) | ✅ Yes | README |
| 71 | StyleTTS2 | TTS | Human-level speech synthesis using style diffusion and adversarial training | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 72 | Tacotron2 | TTS | Industry-classic recurrent sequence-to-sequence Mel-spectrogram generator | No | $0 (CPU-ready) | ✅ Yes | README |
| 73 | Tortoise-TTS | TTS | Studio-quality multi-voice narration, audiobooks, and luxury voice generation | Recommended | $15–$30 (GPU required for speed) | ⚠️ Dedicated recommended | README |
| 74 | VALL-E-X | TTS | Zero-shot cross-lingual speech synthesis and speech-to-speech translation | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 75 | Conformer-CTC | STT | Enterprise call-center ASR combining self-attention with depthwise convolutions | No | $0 (runs on CPU) | ✅ Yes | README |
| 76 | Data2Vec-Audio-Large | STT | Unified multi-modal self-supervised architecture applied to speech recognition | No | $0 (runs on CPU) | ✅ Yes | README |
| 77 | FasterWhisper | STT | Fast audio transcription & voice command parsing (Uzbek & Multi) | Optional | $0–$10 (CTranslate2 on CPU) | ✅ Yes | README |
| 78 | FunASR-Paraformer | STT | Non-autoregressive industrial speech recognition with continuous streaming support | No | $0 (runs on CPU) | ✅ Yes | README |
| 79 | HuBERT-Large | STT | Self-supervised speech representation model with k-means acoustic clustering | No | $0 (runs on CPU) | ✅ Yes | README |
| 80 | Insanely-Fast-Whisper | STT | Batched ultra-fast pipeline inference using Flash Attention 2 and Hugging Face Optimum | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 81 | MMS-1B-All | STT | Meta 1,400-language foundation ASR model for universal speech recognition | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 82 | Moonshine-Base | STT | Low-latency live microphone speech transcription for real-time captions | No | $0 (runs on CPU) | ✅ Yes | README |
| 83 | Moonshine-Tiny | STT | Instant edge streaming speech recognition without 30-second chunk latency | No | $0 (runs on microcontrollers) | ✅ Yes | README |
| 84 | NeMo-Canary-1B | STT | Multitask speech transcription and translation with punctuation and capitalization | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 85 | SeamlessM4T-STT | STT | Multilingual automatic speech recognition and speech-to-text translation across 100+ languages | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 86 | SenseVoice-Small | STT | Ultra-fast speech recognition (<100ms latency) with rich audio event and emotion detection | No | $0 (runs on CPU) | ✅ Yes | README |
| 87 | Sherpa-ONNX-STT | STT | Offline edge embedded speech recognition for Linux, Android, and IoT SBCs | No | $0 (runs on edge / $3 VPS) | ✅ Yes | README |
| 88 | Silero-STT | STT | Enterprise-grade compact models running on single CPU thread (<30MB) with real-time streaming | No | $0 (runs on micro-server) | ✅ Yes | README |
| 89 | Vosk-API-UZ | STT | Lightweight Kaldi-based offline speech recognizer (~50MB model) for mobile and Raspberry Pi | No | $0 (runs on CPU / embedded) | ✅ Yes | README |
| 90 | Wav2Vec2-XLSR-UZ | STT | End-to-end self-supervised acoustic CTC model fine-tuned for Uzbek speech | No | $0 (runs on CPU) | ✅ Yes | README |
| 91 | WavLM-Large | STT | Speech recognition with native background denoising and speaker verification | No | $0 (runs on CPU) | ✅ Yes | README |
| 92 | Whisper-Base | STT | Optimal balance of edge speed and acceptable transcription accuracy for bots | No | $0 (runs on existing server) | ✅ Yes | README |
| 93 | Whisper-Diarization | STT | Multi-speaker meeting transcription with speaker identification ('Who Spoke When') | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 94 | Whisper-Large-v3-Turbo | STT | High-accuracy enterprise transcription at 8x speed (pruned 4-layer decoder) | Optional | $0–$15 (CPU / GPU) | ✅ Yes | README |
| 95 | Whisper-Medium | STT | Production-grade call-center analytics and complex multi-speaker transcription | Recommended | $15–$30 (entry GPU) | ⚠️ Dedicated recommended | README |
| 96 | Whisper-Small-UZ | STT | Domain-fine-tuned Uzbek speech recognition on Common Voice Uzbek corpus | No | $0 (runs on CPU) | ✅ Yes | README |
| 97 | Whisper-Timestamped | STT | Accurate word-level timestamping and subtitle synchronization without hallucinations | No | $0 (runs on CPU) | ✅ Yes | README |
| 98 | Whisper-Tiny | STT | Ultra-lightweight real-time transcription on low-end CPUs and IoT hardware | No | $0 (runs on $4 VPS) | ✅ Yes | README |
| 99 | Whisper.cpp | STT | High-performance pure C/C++ ASR inference with zero Python runtime overhead | No | $0 (runs on bare-metal CPU) | ✅ Yes | README |
| 100 | Zipformer-Transducer | STT | Next-gen Kaldi Zipformer multi-rate transducer with exceptional parameter efficiency | No | $0 (runs on CPU) | ✅ Yes | README |
(100 of 100 models reviewed, scaffolded and verified — 100% COMPLETE)
100-opensource-models-review/
├── README.md # this file — overall index
├── _template/ # template files and rules for adding a new model
├── cv/
│ ├── requirements.txt # unified master dependencies for all 27 CV models
│ ├── bytetrack-mot/
│ ├── depth-anything-v2-small/
│ ├── insightface-arcface/
│ ├── opencv-video-stream/
│ ├── yolo-object-detection/
│ ├── yolo-pose-sleeping/
│ ├── yolo11m/
│ ├── yolo11n/
│ ├── yolo11n-pose/
│ ├── yolo11n-seg/
│ ├── yolo11s/
│ ├── yolov10n/
│ ├── yolov10s/
│ ├── yolov3-tiny/
│ ├── yolov4-tiny/
│ ├── yolov5m/
│ ├── yolov5n/
│ ├── yolov5s/
│ ├── yolov6n/
│ ├── yolov6s/
│ ├── yolov7-tiny/
│ ├── yolov8m/
│ ├── yolov8n/
│ ├── yolov8s/
│ ├── yolov8s-worldv2/
│ ├── yolov9s/
│ ├── yolov9t/
├── llm/
│ ├── DeepSeek-Coder-V2-Lite-Instruct-GGUF/
│ ├── DeepSeek-R1-Distill-Llama-8B-GGUF/
│ ├── DeepSeek-R1-Distill-Qwen-1.5B-GGUF/
│ ├── Gemma-2-2B-Instruct-GGUF/
│ ├── Granite-3.0-2B-Instruct-GGUF/
│ ├── Hermes-3-Llama-3.2-3B-GGUF/
│ ├── Llama-3.1-8B-Instruct-GGUF/
│ ├── Llama-3.2-1B-Instruct-GGUF/
│ ├── Llama-Guard-3-1B-GGUF/
│ ├── Mistral-7B-Instruct-v0.3-GGUF/
│ ├── Moondream2-GGUF/
│ ├── NLLB-200-Distilled-600M/
│ ├── Phi-3.5-mini-instruct-GGUF/
│ ├── Qwen2-VL-2B-Instruct-GGUF/
│ ├── Qwen2.5-1.5B-Instruct-GGUF/
│ ├── Qwen2.5-3B-Instruct-GGUF/
│ ├── Qwen2.5-Coder-1.5B-Instruct-GGUF/
│ ├── SmolLM2-1.7B-Instruct-GGUF/
│ ├── StarCoder2-3B-GGUF/
│ ├── bge-m3/
│ ├── bge-reranker-v2-m3/
├── tts/
│ ├── bark-small/
│ ├── chattts/
│ ├── coqui-tts-vits/
│ ├── coqui-xtts-v2/
│ ├── cosyvoice-300m/
│ ├── e2-tts/
│ ├── espeak-ng/
│ ├── f5-tts/
│ ├── fastspeech2/
│ ├── fish-speech-1.5/
│ ├── glow-tts/
│ ├── matcha-tts/
│ ├── meta-voice-1b/
│ ├── mms-tts-kaz/
│ ├── mms-tts-uzb/
│ ├── mms-tts-uzb-latin/
│ ├── openvoice-v2/
│ ├── parler-tts-mini/
│ ├── piper-tts/
│ ├── seamless-m4t-tts/
│ ├── sherpa-onnx-offline-tts/
│ ├── speecht5-tts/
│ ├── styletts2/
│ ├── tacotron2/
│ ├── tortoise-tts/
│ ├── valle-x/
├── stt/
│ ├── FasterWhisper/
│ ├── conformer-ctc/
│ ├── data2vec-audio-large/
│ ├── funasr-paraformer/
│ ├── hubert-large-ls960/
│ ├── insanely-fast-whisper/
│ ├── mms-1b-all/
│ ├── moonshine-base/
│ ├── moonshine-tiny/
│ ├── nemo-canary-1b/
│ ├── seamless-m4t-stt/
│ ├── sensevoice-small/
│ ├── sherpa-onnx-offline-stt/
│ ├── silero-stt/
│ ├── vosk-api-uz/
│ ├── wav2vec2-large-xlsr-uz/
│ ├── wavlm-large/
│ ├── whisper-base/
│ ├── whisper-cpp/
│ ├── whisper-diarization/
│ ├── whisper-large-v3-turbo/
│ ├── whisper-medium/
│ ├── whisper-small-uz/
│ ├── whisper-timestamped/
│ ├── whisper-tiny/
│ ├── zipformer-transducer/
└── benchmark_scripts/ # shared benchmark scripts per category
Each model folder is self-contained — it has its own Dockerfile, run_benchmarks.py, and README.md.
cv/venv-cv virtual environment (pip install -r cv/requirements.txt) with full local webcam (--source 0), GUI preview, or --headless batch processing.docker compose up <model-name> --build) or locally via Python.Rules and template files for adding a new model — _template/HOW_TO_ADD_A_MODEL.md.