Welcome to TacFlow! If you're here, you probably want to know if your computer can run the platform. Below you'll find the technical specifications broken down into Minimum (it runs, but may be slow) and Recommended (full, fluid experience), with the reason behind each requirement.

Summary Table

Component	️ Minimum	Recommended	Rationale
Operating System	Windows 10 21H2+, macOS 13+, Ubuntu 22.04+	Windows 11 23H2+, macOS 15+, Ubuntu 24.04+	Updated systems ensure compatible drivers and security patches.
CPU	4 cores, 2.0 GHz (Intel i5-8400 / AMD Ryzen 3 3300X)	8+ cores, 3.5+ GHz (Intel i7-12700 / AMD Ryzen 7 5800X)	The runtime runs multiple agents in parallel. Each agent consumes ~1 CPU thread for lightweight inference. For 5+ simultaneous agent workflows, 8 cores prevent bottlenecks.
RAM	8 GB	32 GB	A local LLM (via Ollama/Llama.cpp) needs ~8 GB just to load a 7B 4-bit model. With 16+ agents, browser, and OS, you'll easily hit 12-16 GB.
GPU (VRAM)	4 GB VRAM (NVIDIA GTX 1650 / AMD RX 6400)	12+ GB VRAM (NVIDIA RTX 4070 / AMD RX 7800 XT)	LLMs run 10-50x faster on GPU than CPU. A Llama-3-70B 4-bit needs ~40 GB VRAM. For 7B-13B models, 12 GB is the cost-effectiveness sweet spot.
Storage	20 GB free (SSD)	100+ GB free (NVMe SSD)	Each LLM takes 4-10 GB. Embedding models + cache add up fast. NVMe reduces model loading time by 3-5x vs SATA SSD.
Internet	10 Mbps	100+ Mbps	Model downloads (4-40 GB each), agent updates, and cloud communication require stable bandwidth. High latency (>100ms) degrades API calls.
GPU Drivers	CUDA 12.1+ (NVIDIA) / ROCm 5.7+ (AMD)	CUDA 12.4+ / ROCm 6.0+	Newer drivers have optimized kernels for LLMs (Flash Attention, Tensor Cores). Older versions lose 20-40% inference performance.
Python	3.10+	3.12+	Modern ML libraries (PyTorch, TensorFlow, vLLM) require Python 3.11+. Python 3.12+ brings performance improvements (up to 30% faster loops).
Browser	Chrome 120+, Firefox 121+, Edge 120+	Chrome 126+	TacFlow's UI uses WebGPU and WebAssembly to accelerate local inference. Recent browsers support SharedArrayBuffer and Web Workers.

️ Component Deep Dive

️ Operating System

• Minimum: Windows 10 21H2, macOS 13 (Ventura), Ubuntu 22.04 LTS
• Recommended: Windows 11 23H2, macOS 15 (Sequoia), Ubuntu 24.04 LTS
• Why: Updated operating systems ensure support for GPU accelerators, better memory management, and security. On macOS, older versions lack Metal GPU acceleration. Linux on newer kernels has better performance characteristics.

GPU (Graphics Card)

• Minimum: 4 GB VRAM — Runs 3B-7B models (Phi-3, Llama-3-8B 4-bit) at ~5-10 tokens/s
• Recommended: 12-24 GB VRAM — Runs 13B-70B models (Mistral, Llama-3-70B 4-bit) at ~20-50 tokens/s
• Why: VRAM determines which model fits on the GPU. Larger models are smarter (MMLU: 7B=64%, 70B=86%). Below 8 GB, you're limited to small models or CPU inference (10x slower).

RAM

• Minimum: 8 GB — OS + browser + 1 lightweight agent only
• Recommended: 16 GB — OS + browser + 3-5 agents + 7B model on CPU
• Ideal: 32 GB — Multiple agents + 7B-13B model + cache
• Why: CPU-loaded models keep weights in RAM. A 7B 4-bit model takes ~4 GB in weights alone. Add 2-3 GB inference overhead plus OS and browser — 8 GB fills up quickly.

Storage

• Minimum: 20 GB SSD — 1 small model only
• Recommended: 100+ GB NVMe — Multiple models + datasets
• Why: Models are heavy: Llama-3-8B (~16 GB), Mistral-7B (~14 GB), embeddings (~2 GB each). With 3-4 models, you'll easily exceed 50 GB. NVMe is critical for fast model loading.

How to Check Your Specs

Windows

# Check RAM and CPU
Get-ComputerInfo -Property "OsTotalVisibleMemorySize", "CsProcessors"

# Check GPU VRAM (NVIDIA)
nvidia-smi --query-gpu=memory.total --format=csv,noheader

# Check free disk space
Get-PSDrive -PSProvider FileSystem | Select-Object Name, @{n="Free(GB)";e={[math]::Round($_.Free/1GB,1)}}

macOS

# Check RAM and CPU
system_profiler SPHardwareDataType | grep -E "Memory|Cores|Chip"

# Check disk space
df -h /

# Check GPU (Apple Silicon)
system_profiler SPDisplaysDataType

Linux

# Check RAM and CPU
free -h && nproc

# Check GPU VRAM (NVIDIA)
nvidia-smi --query-gpu=memory.total --format=csv,noheader

# Check disk space
df -h /

Windows Warning — SmartScreen / Unrecognized Software

️	Windows Installation
Issue	When downloading the TacFlow installer, Windows may display: "Windows protected your PC — Microsoft Defender SmartScreen prevented an unrecognized app from starting"
Reason	TacFlow is new software without a digital certificate yet. SmartScreen blocks apps without an established reputation. This is normal and expected.
Solution	Click "More info" → "Run anyway". The installer is safe — verified and signed by the TacFlow team.
Verify	After installing, check the checksum: certutil -hashfile TacFlowSetup.exe SHA256 and compare it with the hash published on the official website.

Tip: If your hardware doesn't meet the minimum requirements, you can still use TacFlow via the cloud — visit app.tacflow.ai and get started immediately without installing anything!

Post created by SupportDev — Questions? Ask below!