MMOOZZGG.com

Hardware

Run LLM on MacBook: what fits, how fast it is, and when a laptop is not enough

3 min read8 sources

To run an LLM on a MacBook is now a simple way into local AI: up to 128 GB of unified memory, one-click programs and silence. Below is what runs at each memory size, measured speed, the current price and the limits beyond which you need a Mac Studio or a server.

On this page · 5 sections

What runs at each memory size

16–24 GB: models of 4–8 billion parameters in 4-bit quantization, an assistant for text and code, 5 GB per model. 36–48 GB: models of 20–35 billion, including a large Russian bank’s open 32-billion model, and at 48 GB a vision model next to it. 64 GB: 70-billion models in 4 bits with a short context. 128 GB: 70 billion with a long context, several models at once, and models with few active parameters at 35–120 billion. The personal agent of a bank’s IT director lives on a Mac, on a 35-billion model with 3 billion active parameters, at 70 tokens per second.

Speed

By the llama.cpp community table, on a 7-billion model in 4 bits: M4 Pro 50.7 tokens per second, M4 Max 83.1, M5 Max 119.9; prompt processing is 3,220 tokens per second on the M5 Max and 886 on the M4 Max. For a chat or an agent on a laptop, this is more than enough; the laptop’s limits are heat and battery under hours of load.

Price as of 29 September 2026

MacBook Pro 14 M5 Max, 128 GB of memory, 2 TB: $8,100 at a Russian seller, in stock. Configurations with 48 and 64 GB cost less; check the seller’s prices on the day you buy. For comparison: one 32 GB graphics card costs $9,000 without a computer (30 September 2026), and a Mac Studio M4 Max with 128 GB costs $11,000.

Software

LM Studio and Ollama install a model in minutes and give you a local API that works with common libraries; agents connect to it as they would to the cloud. This is how the personal agent from the bank case and a cloud provider’s agent work. Several users need a server and a batch serving engine.

Limits of a laptop

One user, not a service for a department. Long documents process more slowly than on a server card. The laptop leaves with the employee, together with the model and the data, so disk encryption is a must. For a team, the next step is a Mac Studio or a server with a graphics card.

Questions and answers

Is a MacBook Air enough?

For models of 4–8 billion parameters, yes, with 16–24 GB of memory. For 30 billion and above, you need a Pro with 36 GB or more.

Do you need the internet?

Only to download the model. After that everything works offline, which is the whole point of local AI.

Sources and fact-checking

Checked by the editors on 29 September 2026.

  1. apple-store.net.ru, MacBook Pro 14 M5 Max 128 GB 2 TB, 29 September 2026 (apple-store.net.ru)
  2. llama.cpp, performance on Apple Silicon, table updated 25 August 2026 (github.com)
  3. Regard, GeForce RTX 5090, prices as of 29 September 2026 (regard.ru)
  4. store-apple.msk.ru, Mac Studio M4 Max 128 GB, 29 September 2026 (store-apple.msk.ru)
  5. Digital Platforms, case No. 2447, October 2025: T-Bank (platforms.su)
  6. Digital Platforms, case No. 3855, 2 September 2026: T-Technologies (platforms.su)
  7. Digital Platforms, case No. 4067, 2 September 2026: Alfa-Bank, IT director (platforms.su)
  8. iXBT, T-Bank released T-Pro 2.0 (32B, Apache 2.0), 18 July 2025 (ixbt.com)

Found a mistake? Tell us

Publications

Cases on this topic: 3

  • Market case · Russia

    A large Russian bank in-house LLMs solve 45% of requests; support costs estimated down at least 25% a year

    The bank put its own family of language models into its chatbot for LLM customer support: about 45% of customer requests are solved without an operator. In October 2025 it launched AI copilots for first-line operators, and more than 90% of support staff use them every day. The models run inside the bank’s network perimeter, so data does not leave the bank.

    Banks and finance

    source: platforms.su

    Read the case
  • Market case · Russia

    A large Russian financial group AI financial assistant on in-house language models, beta since 20 July 2026

    The group built an AI financial assistant for tasks linked to money and shopping, on its own family of large language models. To avoid made-up answers, the system relies on verified data sources, not only on what the model has learned. The beta was presented on 20 July 2026, and the full launch is planned for the autumn.

    Banks and finance

    2 September 2026 · source: platforms.su

    Read the case
  • Market case · Russia

    Alfa-Bank, IT head local agent on the Qwen 3.6–35B-A3B model: about 70 tokens per second on a Mac

    A local AI agent on Mac does autonomous work on a model with 35 billion parameters, about 3 billion of them active per token. It runs through LM Studio at about 70 tokens per second. The agent first ran in Docker on the Mac, then moved to a virtual machine, and the author calls that move a turning point for ease of use.

    Banks and finance

    2 September 2026 · source: platforms.su

    Read the case
Contact us