Running a local LLM for video notes on your Mac

ScribeShot can generate notes, chapters, summaries, and screenshot analysis without sending anything to an API. This page covers which model to pick, what it actually downloads, how much RAM you need, and the cases where a cloud key is still the better answer.

What “local” means here

The model runs as a process on your Mac against a quantised GGUF file on your own disk. There is no API key, no account, and no request leaving the machine for note generation. Once the model file is downloaded, the feature works with the network off.

Two consequences worth being clear about. First, the download is measured in gigabytes, not megabytes — this is the real cost of “no subscription”, paid once in disk space rather than monthly. Second, generation speed is bounded by your hardware, so an older Mac will be slower than an API call, not faster.

The three models

Each entry ships a vision projector alongside the language weights, because screenshot analysis runs through the same local model. Both files are fetched together, so plan for the combined size:

Files live under ~/Library/Application Support/ScribeShot/models/llm. They are ordinary GGUF files; deleting one frees the space and the app will offer the download again.

Why nothing smaller than about 4B

This is the most useful thing on this page if you are evaluating local-model tools generally. Smaller models are widely available and tempting — Llama 3.2 1B and 3B, Phi-3 and Phi-4 Mini, Qwen 2.5 1.5B and 3B, Gemma 2 2B, SmolLM3 3B. ScribeShot deliberately does not ship any of them.

The reason is not raw quality, it is structured output reliability. The notes, chapters, and summary pipeline needs valid JSON matching a schema on every call. Sub-4B models produced inconsistent structured JSON often enough that the pipeline stayed unreliable even with retry-and-feedback logic. A model that writes decent prose but breaks its own schema one time in six is worse than useless in an automated chain, because the failure is silent and downstream.

Gemma 4 E2B is the single exception, kept because it clears that bar despite its size. If you are assessing another tool that advertises a 1B or 3B local model for structured extraction, this is the specific thing worth testing before trusting it.

Quantisation, briefly

All three models ship at Q4_K_M, a 4-bit quantisation. This is the standard practical trade: roughly a quarter of the memory of full precision for a quality loss that is minor on summarisation work. It is why a 31-billion-parameter model fits in 19 GB on disk rather than well over 60.

When to use OpenAI instead

Local is the default and the right choice for most work, but a key genuinely wins in some cases:

Both are supported and the choice is a setting, not a purchase decision — you can switch per use. Supported OpenAI models run from gpt-5-nano for cheap throughput to gpt-5.1 for the strongest reasoning, with gpt-5-mini as the default. You supply your own key and pay OpenAI directly; ScribeShot takes no cut and adds no markup. The key is stored in your Mac's system keychain, not on a server.

Common questions

Do I need an OpenAI API key?

No. ScribeShot ships a local LLM engine that runs Gemma 4 on your Mac with no key and no account. An OpenAI key is optional and useful mainly for very long videos or on machines where you would rather not spend the RAM.

How much disk space does the local model need?

About 2 GB for Gemma 4 E2B, 3.6 GB for the E4B default, and 19 GB for the 31B flagship. Each figure includes the vision projector file that is downloaded alongside the language weights.

How much RAM do I need?

The E2B and E4B models are intended for typical Macs, with E4B as the default. The 31B model requires 24 GB or more of RAM.

Why does ScribeShot not offer smaller local models like Phi-3 Mini or Llama 3.2 3B?

The notes and chapters pipeline requires valid schema-conforming JSON on every call. Models below roughly 4B parameters produced inconsistent structured output in testing, which made the pipeline unreliable even with retries. Gemma 4 E2B is the one exception that met the bar.

Does the local model handle screenshots too?

Yes. Screenshot analysis runs through the same local model using the vision projector downloaded with it, which is why the download includes that extra file.