Local LLM model fit
Devstral 2 123B is a 123B Devstral model. This page estimates Q4 VRAM fit, Ollama command, context planning, and fallback choices for common local AI GPUs.
High-end local software engineering agents, large-repository navigation, and tool-heavy coding workflows on server-class memory. Weakness: 75GB Q4_K_M footprint is not realistic for single consumer GPUs; expect server GPU, multi-GPU, unified memory, or cloud fallback.
| Hardware | Examples | Clean capacity | Q4 need | Status | Calculator |
|---|---|---|---|---|---|
| 6 GB VRAM entry GPU | GTX 1660, RTX 2060 6GB | 4.5 GB clean VRAM | 75 GB | Too large | Open calculator |
| 8 GB VRAM mainstream GPU | RTX 3060 Ti, RTX 4060, RTX 3070 | 6.5 GB clean VRAM | 75 GB | Too large | Open calculator |
| 10 GB VRAM older high-end GPU | RTX 3080 10GB | 8.5 GB clean VRAM | 75 GB | Too large | Open calculator |
| 12 GB VRAM local agent GPU | RTX 3060 12GB, RTX 4070, RTX 5070 | 10.5 GB clean VRAM | 75 GB | Too large | Open calculator |
| 16 GB VRAM creator GPU | RTX 4060 Ti 16GB, RTX 4080, RTX 5070 Ti, RTX 5080 | 14.5 GB clean VRAM | 75 GB | RAM offload | Open calculator |
| 24 GB VRAM homelab workstation | RTX 3090, RTX 4090 | 22.5 GB clean VRAM | 75 GB | RAM offload | Open calculator |
| 32 GB VRAM Blackwell workstation | RTX 5090 | 30.5 GB clean VRAM | 75 GB | RAM offload | Open calculator |
| 48 GB VRAM workstation | RTX A6000, L40S 48GB | 46.5 GB clean VRAM | 75 GB | RAM offload | Open calculator |
| Apple Silicon 32 GB unified memory | M2 Max 32GB, M3 Max 36GB | 26 GB unified | 75 GB | Too large | Open calculator |
| Apple Silicon 256 GB unified memory | Mac Studio M3 Ultra 256GB, Mac Studio M4 Ultra 256GB | 250 GB unified | 75 GB | Runs locally | Open calculator |
| Quantization | Estimated memory | Use case |
|---|---|---|
| Q4 / 4-bit | 75 GB | Default local inference balance |
| Q5 / 5-bit | 93.8 GB | Better quality, more VRAM |
| Q8 / 8-bit | 150 GB | High quality, much more VRAM |
| FP16 / 16-bit | 300 GB | Mostly workstation/server use |
This is a practical planning estimate, not a benchmark. Real memory use changes with backend, context length, KV cache, quantization file, drivers, and offloading settings.
120B · Large local reasoning servers, heavy agent orchestration, and high-end homelab inference
80B · High-end local coding agents, repository-scale code edits, and tool-calling development workflows
80B · High-end local reasoning, long-context planning, and tool-agent workloads when 48GB+ memory is available
70B · High-quality local chat and reasoning on workstation-class hardware