TokenAssemble

Local AI workflow

Private chat server

A self-hosted ChatGPT-style app for your household or team — nothing leaves the box. The stack: Open WebUI, LibreChat with Ollama serving 3 recommended local models — hardware readiness graded by the same engine as the feasibility checker.

The stack

Hardware readiness

Each cell is the engine's verdict for the best clean quant of that model on Ollama at 8K context — the same math as the checker, tiers not numbers.

Hardwarellama-3.1-8b-instructqwen2.5-14b-instructllama-3.3-70b-instruct
NVIDIA GeForce RTX 4090Fits · A+ · Q8_0interactiveFits · B · Q8_0interactiveTight · C · Q4_K_Musable · offloadbeta
NVIDIA GeForce RTX 5090Fits · A+ · Q8_0interactiveFits · A+ · Q8_0interactiveTight · C · Q4_K_Musable · offloadbeta
NVIDIA GeForce RTX 4080Fits · A+ · Q8_0interactiveFits · B · Q5_K_MinteractiveTight · C · Q4_K_Mpainful · offloadbeta
NVIDIA GeForce RTX 4070Fits · B · Q8_0interactiveTight · C · Q4_K_MinteractiveTight · C · Q4_K_Mpainful · offloadbeta
NVIDIA GeForce RTX 4060Fits · B · Q4_K_MinteractiveTight · C · Q4_K_Musable · offloadbetaTight · C · Q4_K_Mpainful · offloadbeta
NVIDIA GeForce RTX 3090Fits · A+ · Q8_0interactiveFits · B · Q8_0interactiveTight · C · Q4_K_Mpainful · offloadbeta
NVIDIA GeForce RTX 3060 12 GBFits · B · Q8_0interactiveTight · C · Q4_K_MinteractiveTight · C · Q4_K_Mpainful · offloadbeta
AMD Radeon RX 7900 XTXFits · A+ · Q8_0interactivebetaFits · B · Q8_0interactivebetaTight · C · Q4_K_Mpainful · offloadbeta
Intel Arc B580Fits · B · Q8_0interactivebetaTight · C · Q4_K_MinteractivebetaTight · C · Q4_K_Mpainful · offloadbeta
NVIDIA DGX SparkFits · A+ · Q8_0interactivebetaFits · A+ · Q8_0usablebetaFits · A+ · Q8_0painfulbeta
Mac mini M4 24GBFits · A+ · Q8_0painfulFits · B · Q6_KpainfulWon't fit
Mac Studio M3 Ultra 96GBFits · A+ · Q8_0interactiveFits · A+ · Q8_0interactiveFits · B · Q6_Kpainful
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBFits · A+ · Q8_0usablebetaFits · A+ · Q8_0usablebetaFits · B · Q8_0painfulbeta

Different hardware? Check your exact machine against any model in this stack.

Check your machine →Don't have the hardware? Build a stack for your budget →

Template picks are editorial; every verdict, grade and tier above comes from the engine over sourced data — how the math works.