← All projects

Local AI Lab

Claims about local models, checked against my own VRAM

Status Personal hardware. Ongoing.


Why bother

Every claim about what a local model can do is worth exactly as much as the hardware budget of the person making it. Benchmarks are run on cards nobody has at home, and the interesting question is not whether a model is good but whether it is good enough on twelve gigabytes.


The setup

An RTX 3060 with 12 GB of VRAM, Ollama and LM Studio, and a fixed set of agentic development tasks to run everything against so the comparison means something.

Models under evaluation include Qwen, Gemma, Llama 3.1, DeepSeek R1 and Mistral Nemo.


What it actually tells me

Which development workflows survive without a frontier model and which quietly fall apart. Tool calling and MCP integrations are where the gap shows first: a model that writes acceptable code will still lose the thread across a multi-step agentic loop long before it runs out of context.

It is also where I find out what a workflow costs when the marginal token is free, which is the other half of the question TokenMeter asks.


Stack

OllamaLM StudioMCPClaude CodeRTX 3060

Written up on the lab