# Skill: Assess hardware requirements for running local AI models (VRAM, RAM, Apple Silicon)

> Apply this skill when a user needs to know whether their computer can run a specific local model, or when recommending hardware for local AI inference.

Source: sodigi-learn -- local-agents/zhelezo-dlya-ii -- https://sodigi.io/learn/local-agents/zhelezo-dlya-ii

## When to use
- User wants to run a local model and asks if their machine is capable.
- User reports the model is slow, freezing, or refusing to start.
- User is choosing between running on GPU vs CPU vs Apple Silicon.
- User is considering buying hardware specifically for local AI workloads.

## Core rules
- VRAM (video card memory) is the single most important factor -- not CPU clock, not core count.
- A model runs fast only when it fits entirely in VRAM; any overflow into RAM causes a major speed drop.
- RAM is the fallback: the tail of the model can spill into it, but speed falls several times over.
- CPU-only (no GPU) is a last resort, workable only for tiny models and tests.
- Apple Silicon (M-series chips) has unified memory: the GPU and CPU share the same pool, so a Mac with 32-64 GB effectively has that much VRAM -- a significant advantage.
- On Apple Silicon the memory is soldered; you cannot upgrade it after purchase.
- Always leave a buffer -- never pick a model that fills VRAM to the edge; context grows during a conversation.
- Quantization (4-bit compression) shrinks a model roughly fourfold with minimal quality loss; it is the primary rescue tool for limited hardware.

## Procedure
1. Ask the user for their GPU model and its VRAM size in GB. If no discrete GPU, ask total RAM.
2. For Apple Silicon Macs, ask the total unified memory (the spec that counts).
3. Apply the rough 4-bit quantized ballpark:
   - 7B/8B model: ~5-6 GB VRAM
   - 13B model: ~8-10 GB VRAM
   - 70B model: ~40 GB+ (requires professional card, dual-GPU, or high-memory Mac)
4. Check whether the target model fits with at least 2 GB to spare for context.
5. If it does not fit: recommend either a smaller model size or a lower quantization level (e.g., Q4 instead of Q8).
6. If the user has no GPU: recommend a 7B or smaller model in Q4 form, warn about slow CPU inference.
7. Remind the user to pick quantized (GGUF/Q4) variants, not raw uncompressed weights.

## Ready-to-use prompt
```
Help me figure out what local AI model this computer can run.

Hardware profile:
- GPU: [model name, if any]
- GPU VRAM: [X GB, if applicable]
- CPU RAM: [X GB]
- Platform: [Windows / Linux / Mac with M-chip, specify chip generation and total RAM]

Please:
1. State whether a 7B, 13B, or 70B quantized model (Q4) fits in VRAM with a buffer.
2. If none fit in VRAM, state what fits in RAM as an offload fallback and warn about speed.
3. Recommend the largest model size that fits comfortably (VRAM with ~2 GB spare).
4. Name the quantization level to use (Q4 for tight hardware, Q5/Q8 if memory is plentiful).
5. Flag if the machine is CPU-only and set realistic speed expectations.
```

## Pitfalls
- Looking at the GPU model name or clock speed instead of VRAM size -- VRAM is the only number that matters.
- Recommending a model that fills VRAM exactly; context length during a conversation will overflow it.
- Forgetting that Apple Silicon unified memory is the GPU memory too -- do not treat it as ordinary RAM.
- Suggesting the user download an uncompressed model when a Q4 GGUF version is available.
- Blaming a "slow computer" when the real issue is a model two sizes too large for the hardware.
- Recommending no model at all for modest hardware; a small Q4 model almost always runs acceptably.