Skip to main contentThe Local Bench

Use case 8 of 13

A patient maths tutor

Step-by-step help with homework and exam practice on a laptop, with no accounts, no ads and no data collection on children.

A notebook, pencil, protractor and wooden blocks with one blue cube beside a tablet
Image: generated with Grok Imagine

The real example

Phi-4-mini-reasoning, built for embedded tutoring

Microsoft describes Phi-4-mini-reasoning, a 3.8B model, as potentially suitable for educational applications, embedded tutoring and lightweight deployment on edge or mobile systems. Its model card reports 94.6% on MATH-500 and 57.5% on AIME, well above the base Phi-4-mini's 71.8% and 10.0%. Larger reasoning models go further: Microsoft reports 75.3% on AIME 2024 for the 14B Phi-4-reasoning.

Sources: Phi-4-mini-reasoning model card · Phi-4-reasoning model card

Why run it locally

Children's data deserves the strictest privacy, and a tutor that runs offline works on the bus. Both models are MIT-licensed.

Recommended stack

Three tiers built from our builds. Weights-in-memory is our arithmetic (parameters × bits), checked against each device's usable memory; it excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Budget

Value 16 GB

$1,173Sample

16 GB on GeForce RTX 5060 Ti 16 GB

Weights ≈ 2.3 GB of 15.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Also runs on many laptops; the build gives quick, long reasoning traces.

Runtime: LM Studio or Ollama

Setup outline

  1. Install LM Studio or Open WebUI with a local model server.
  2. Write a tutoring system prompt that asks for hints first and full solutions only on request.
  3. For worksheets, add a vision model so students can photograph problems.
  4. Have the student check each step. Reasoning models can be confidently wrong.

Limits

  • Benchmark scores are vendor-reported, and AIME is competition maths, not pedagogy.
  • Long reasoning traces are slow on small hardware; a 3.8B model can still take tens of seconds per answer.
  • A tutor that hands out answers can undermine learning. Configure it to coach.

Next: Answering the phone for a small business

Build prices come from our build datasheets: GPUs and devices are sourced retail or sold prices; other parts are sample prices until sourced, and the flag on each total says which. Model links go to our datasheets, where each parameter count and licence was checked against the Hugging Face API.