Monday, October 26, 2026 · Edition updated 6:00 AM ETMON, OCT 26 editionNext edition in Next in --:--:--
Intelligence Verdict
Get the weekly brief
Run AI on your own computer

AI that runs on your own machine.

Open models you can download and use offline: how big each one is, how much memory it takes, how fast it runs on Macs and graphics cards, what the machines cost, and the free apps that run them.

  • 31open models, sized from their download pages
  • 16 GBof memory, which OpenAI says is enough for gpt-oss-20b
  • 300tokens a second on a RTX 5090, for a small model
What can my computer run?

Pick your machine.

On a Mac, and on PCs with shared memory, models use the computer's main memory. On a PC with a graphics card, they run fastest when the whole model fits on the card.

Memory in the computer

15 run comfortably

The file takes up to three quarters of 16 GB.

  • gpt-oss 20bOpenAI
    12 GBMXFP4
  • Ministral 3 14BMistral AI · France
    9.1 GBQ4_K_M
  • Phi-4Microsoft
    9.1 GBQ4_K_M
  • Gemma 4 12BGoogle DeepMind
    7.2 GBQAT (4-bit)
  • Qwen3.5 9BAlibaba · China
    6.6 GBQ4_K_M
  • Ministral 3 8BMistral AI · France
    6 GBQ4_K_M
  • Gemma 4 E4BGoogle DeepMind
    5.9 GBLM Studio default
  • Granite 4.2 8BIBM
    5.3 GBQ4_K_M
  • LFM2.5 8B-A1BLiquid AI
    5.2 GBOllama default
  • Olmo 3 7B InstructAllen Institute for AI
    4.5 GBQ4_K_M
  • Gemma 4 E2BGoogle DeepMind
    4.2 GBLM Studio default
  • Nemotron 3 Nano 4BNVIDIA
    2.8 GBQ4_K_M
  • Phi-4-mini-instructMicrosoft
    2.5 GBQ4_K_M
  • Ministral 3 3BMistral AI · France
    2 GBLM Studio default
  • Qwen3.5 2BAlibaba · China
    1.9 GBQ4_K_M

1 tight fit

Fits, with little room left. Close other apps and keep conversations short.

  • Gemma 4 26B A4BGoogle DeepMind
    15.6 GBLM Studio default

15 too big

The file alone is larger than 16 GB. Smallest first.

  • Qwen3.6 27BAlibaba · China
    16.1 GBLM Studio default
  • Qwen3.8 27BAlibaba · China
    16.1 GBLM Studio default
  • Muse Glimmer 30BMeta
    18 GBQ4_K_M
  • Granite 4.2 30BIBM
    18 GBQ4_K_M
  • Gemma 4 31BGoogle DeepMind
    19 GBQAT (4-bit)
  • Qwen3.6 35B-A3BAlibaba · China
    20.4 GBLM Studio default

and 9 larger, up to 1.51 TB

Machines that come with 16 GB or more
  • Mac mini (M6)16 GB from $899
  • MacBook Air 13-inch (M5)16 GB from $1,099
  • MacBook Air 15-inch (M5)16 GB from $1,299
  • MacBook Pro 14-inch (M5)16 GB from $1,599
  • Mac mini (M5 Pro)24 GB from $1,699
  • MacBook Pro 14-inch (M5 Pro / M5 Max)24 GB from $2,199
  • MacBook Pro 16-inch (M5 Pro / M5 Max)24 GB from $2,699
  • Framework Desktop (AMD Ryzen AI Max / Max+)32 GB from $1,269
  • Mac Studio (M5 Max / M5 Ultra)36 GB from $2,499
  • NVIDIA DGX Spark64 GB from $4,999

We sort by each model's usual download: its smallest 4-bit file. A file up to three quarters of the memory counts as comfortable; that is our rule of thumb, not the makers'. Their own guidance is below.

In the makers' words

How much memory you need.

What the apps and machine makers say. The model file has to fit, and a longer conversation takes more memory on top.

“(8GB RAM for 3B models, 16GB for 7B, 32GB for 13B)”
Jan on GitHub (README)
“16GB+ RAM recommended.”
LM Studio docs: System requirements
“You may still be able to use LM Studio on 8GB Macs, but stick to smaller models and modest context sizes.”
LM Studio docs: System requirements
“at least 4GB of dedicated VRAM is recommended.”
LM Studio docs: System requirements
“Setting a larger context length will increase the amount of memory required to run a model.”
Ollama docs: Context length
“< 24 GiB VRAM: 4k context”
Ollama docs: Context length
“For best performance, use the maximum context length for a model, and avoid offloading the model to CPU.”
Ollama docs: Context length
“Test, validate, and inference with AI models up to 200 billion parameters”
NVIDIA DGX Spark product page
“Run 300B-class models at 4-bit quantization with headroom to spare for extra context length.”
Framework Desktop product page
Ollama's default conversation length, by graphics memory
  • Under 24 GiB VRAM: 4k context
  • 24-48 GiB VRAM: 32k context
  • At least 48 GiB VRAM: 256k context
Ollama docs: Context length
The open models

Every model, by size.

Smallest first. Each download size is the file as its page lists it, read Oct 20, 2026. Fewer bits per number (4-bit rather than 8- or 16-bit) makes a smaller file, which llama.cpp offers “for faster inference and reduced memory use”.

Gemma 4 E2B

Google DeepMind · US

4.2 GBLM Studio default
Size
5.1B (with embeddings), 2.3B effective active
License
Apache 2.0
  • QAT (4-bit)4.3 GB
  • Q4_K_M4.6 GB
  • Q8_06 GB
  • BF1610 GB
  • LM Studio default4.2 GB

“Smaller models are specifically designed for efficient local execution on laptops and mobile devices.” Google DeepMind

LM Studio model page: Gemma 4

LFM2.5 8B-A1B

Liquid AI · US

5.2 GBOllama default (quantization not shown)
Size
8.3B, 1.5B active
License
lfm1.0 (as listed on Hugging Face)
  • Ollama default (quantization not shown)5.2 GB

Ollama library: lfm2.5

Gemma 4 E4B

Google DeepMind · US

5.9 GBLM Studio default
Size
8B (with embeddings), 4.5B effective active
License
Apache 2.0
  • QAT (4-bit)6.1 GB
  • Q4_K_M6.6 GB
  • Q8_09 GB
  • BF1616 GB
  • LM Studio default5.9 GB

“deployment scenarios from mobile and edge devices (E2B, E4B) to consumer GPUs and workstations” Google DeepMind

LM Studio model page: Gemma 4

Qwen3.5 9B

Alibaba (Qwen team) · China

6.6 GBQ4_K_M
Size
9B
License
Apache 2.0
  • Q4_K_M6.6 GB
  • Q8_010 GB
  • BF1619 GB
  • LM Studio default7 GB

“If you encounter out-of-memory (OOM) errors, consider reducing the context window.” Alibaba

Ollama library: qwen3.5 tags

Gemma 4 26B A4B

Google DeepMind · US

15.6 GBLM Studio default
Size
25.2B, 3.8B active
License
Apache 2.0
  • QAT (4-bit)16 GB
  • Q4_K_M18 GB
  • Q8_028 GB
  • BF1652 GB
  • LM Studio default15.6 GB

“Their diverse sizes make them deployable in environments ranging from high-end phones to laptops and servers” Google DeepMind

LM Studio model page: Gemma 4

Qwen3.6 27B

Alibaba (Qwen team) · China

16.1 GBLM Studio default
Size
27B
License
Apache 2.0
  • Q4_K_M17 GB
  • Q8_030 GB
  • BF1656 GB
  • LM Studio default16.1 GB

“If you encounter out-of-memory (OOM) errors, consider reducing the context window.” Alibaba

LM Studio model page: Qwen3.6

Qwen3.8 27B

Alibaba (Qwen team) · China

16.1 GBLM Studio default
Size
27B
License
Apache 2.0
  • Q4_K_M18 GB
  • Q8_030 GB
  • BF1656 GB
  • LM Studio default16.1 GB

“Inference efficiency and throughput vary significantly across frameworks.” Alibaba

LM Studio model page: Qwen3.8

Kimi K3

Moonshot AI · China

1.51 TB4-bit (Unsloth dynamic)
Size
2.8T, 104B active
License
Kimi K3 License
  • 1-bit (Unsloth dynamic)466 GB
  • 2-bit (Unsloth dynamic)861 GB
  • 4-bit (Unsloth dynamic)1.51 TB

Hugging Face: unsloth/Kimi-K3-GGUF

Models with “active” parameters are mixtures of experts: only part of the model works on each word, but the whole file still has to fit in memory. The biggest Chinese models (DeepSeek, GLM, Kimi) are open, but their files run from about 100 GB to over 1.5 TB; the sizes shown come from Unsloth, a third party. Country is where each company is based.

How fast it runs

Speed, measured.

The same small model (Llama 2 7B, at 4 bits) on different chips, in tokens a second as it writes. A token is a word or part of a word. These are community-submitted results from the llama.cpp project's own benchmark threads.

Apple M4 and M3 chips
  • M3 Ultra (80-core GPU, 800GB/s)92
  • M4 Max (40-core GPU, 546GB/s)83
  • M4 Pro (20-core GPU, 273GB/s)51
  • M4 (10-core GPU, 120GB/s)24

llama.cpp GitHub discussion #4167

Apple M5 chips
  • M5 Ultra (80-core GPU, 1228GB/s)179
  • M5 Max (40-core GPU, 614GB/s)120
  • M5 (10-core GPU, 154GB/s)32

Run on a newer build of the software; the thread says these aren't directly comparable with the rows above.

llama.cpp GitHub discussion #4167

NVIDIA graphics cards
  • RTX 5090300
  • RTX 4090189
  • RTX 5080185
  • RTX 5070 Ti182
  • RTX 5070128
  • RTX 5060 Ti93

llama.cpp GitHub discussion #15013

NVIDIA DGX Spark, current models
  • gpt-oss 20BMXFP4, 11.27 GiB83
  • gpt-oss 120BMXFP4, 59.02 GiB59
  • Gemma 3 4BQ4_0, 2.35 GiB81
  • Qwen3 Coder 30B-A3BQ8_0, 30.25 GiB61

Measured Feb 5, 2026. llama.cpp benches

Bigger models run slower on the same machine; these show how chips compare, not how fast every model runs. The NVIDIA cards ran with a speed-up setting (flash attention) on. Each row's full setup is in the thread it links to.

The machines

What they cost.

The memory each machine can be ordered with, and the maker's starting price, which buys the smallest memory. More memory costs more; makers' stores show those prices only as you configure.

Macs

Mac mini (M6)

  • 16 GB
  • 24 GB
  • 32 GB

$899 with 16 GB

Apple Newsroom, Aug 25, 2026 · company-reported

Mac mini (M5 Pro)

  • 24 GB
  • 48 GB
  • 64 GB

$1,699 with 24 GB

Apple Newsroom, Aug 25, 2026 · company-reported

MacBook Air 13-inch (M5)

  • 16 GB
  • 24 GB
  • 32 GB

$1,099 with 16 GB

Apple Newsroom, Mar 3, 2026 · company-reported

MacBook Air 15-inch (M5)

  • 16 GB
  • 24 GB
  • 32 GB

$1,299 with 16 GB

Apple Newsroom, Mar 3, 2026 · company-reported

MacBook Pro 14-inch (M5)

  • 16 GB
  • 24 GB
  • 32 GB

$1,599 with 16 GB

Apple Newsroom, Oct 15, 2025 · company-reported

MacBook Pro 14-inch (M5 Pro / M5 Max)

  • 24 GB
  • 36 GB
  • 48 GB
  • 64 GB
  • 128 GB

$2,199 with 24 GB$3,599 with 36 GB

Apple Newsroom, Mar 3, 2026 · company-reported

MacBook Pro 16-inch (M5 Pro / M5 Max)

  • 24 GB
  • 36 GB
  • 48 GB
  • 64 GB
  • 128 GB

$2,699 with 24 GB$3,899 with 36 GB

Apple Newsroom, Mar 3, 2026 · company-reported

Mac Studio (M5 Max / M5 Ultra)

  • 36 GB
  • 48 GB
  • 64 GB
  • 96 GB
  • 128 GB
  • 256 GB
  • 512 GB

$2,499 with 36 GB$5,499 with 96 GB

Apple Newsroom, Aug 25, 2026 · company-reported

PCs and mini PCs

NVIDIA DGX Spark

  • 64 GB
  • 128 GB

$4,999 with 64 GB$6,950 with 128 GB

NVIDIA Marketplace, Oct 20, 2026 · company-reported

Framework Desktop (AMD Ryzen AI Max / Max+)

  • 32 GB
  • 64 GB
  • 128 GB

$1,269 with 32 GB$1,959 with 64 GB$3,449 with 128 GB

Framework, Oct 20, 2026 · company-reported

Copilot+ PC (Windows laptop with an NPU)

No memory minimum on Microsoft's pages; needs a chip with an NPU of 40+ TOPS.

Microsoft Learn, Nov 17, 2025 · company-reported

Graphics cards

GeForce RTX 5090

  • 32 GB

$1,999 with 32 GB

NVIDIA Newsroom, Jan 6, 2025 · launch price

GeForce RTX 5080

  • 16 GB

$999 with 16 GB

NVIDIA Newsroom, Jan 6, 2025 · launch price

GeForce RTX 5070 Ti

  • 16 GB

$749 with 16 GB

NVIDIA Newsroom, Jan 6, 2025 · launch price

GeForce RTX 5070

  • 12 GB

$549 with 12 GB

NVIDIA Newsroom, Jan 6, 2025 · launch price

GeForce RTX 5060 Ti

  • 8 GB
  • 16 GB

$379 with 8 GB$429 with 16 GB

NVIDIA Newsroom, Apr 15, 2025 · launch price

GeForce RTX 5060

  • 8 GB

$299 with 8 GB

NVIDIA Newsroom, Apr 15, 2025 · launch price

Graphics card prices are NVIDIA's launch prices from 2025; store prices now may differ. A graphics card also needs a PC to go in. The Framework Desktop showed out of stock when we read it, and NVIDIA says partners sell the 64 GB DGX Spark from October 23.

The apps

Free apps that run them.

Each downloads open models and runs them on your machine. Most are free; some add paid online models.

Ollama

Ollama

Downloads and runs open models with one command or its app; also offers cloud-hosted models. Docs: 'Ollama runs locally. We don't see your prompts or data when you run locally.'

Price: Free (MIT-licensed); Ollama also offers cloud-hosted models

  • macOS 14 Sonoma or newer
  • Windows 10 22H2 or newer
  • Linux
  • Docker

Ollama on GitHub · company-reported

LM Studio

LM Studio

Desktop app to search, download and chat with local models; 'LM Studio Bionic combines local AI features with optional cloud inference.'

Price: Free tier $0 (Run local LLMs on your machine); paid Bionic+ $20/month and Pro $100/month add cloud models

  • macOS 14 or newer on Apple Silicon
  • Windows x64
  • Linux x64 and ARM64

LM Studio pricing page and system requirements · company-reported

llama.cpp

ggml-org (open-source project)

'The main goal of `llama.cpp` is to enable LLM (and VLM) inference with minimal setup and state-of-the-art performance'. Supports NVIDIA CUDA, AMD HIP, Vulkan, Intel SYCL and more.

Price: Free

  • macOS
  • Windows
  • Docker
  • Android
  • Pre-built binaries or build from source

llama.cpp on GitHub · company-reported

Jan

Jan (janhq)

Jan is an open source alternative to ChatGPT that runs 100% offline on your computer.

Price: Free (Free & Open source)

  • Windows 10 or later
  • macOS 13.6 or later
  • Linux

Jan on GitHub · company-reported

GPT4All

Nomic AI

'GPT4All runs large language models (LLMs) privately on everyday desktops & laptops.' Latest dated release on the GitHub page is v3.0.0 (July 2, 2024), so it may be less actively updated.

Price: Free

  • Windows x64 and ARM
  • macOS 12.6 Monterey or newer
  • Linux x86-64

GPT4All on GitHub · company-reported

Apple Intelligence on-device model

Apple

Apple describes 'a compact, approximately 3-billion-parameter model' running on device, compressed 'to 2 bits per weight (bpw)'; 'App developers can take advantage of AI inference that is free of cost and accessible with just a few lines of code.'

Price: Free to app developers

  • Apple devices with Apple Intelligence

Apple Machine Learning Research · company-reported

Foundry Local

Microsoft

'Foundry Local is an end-to-end local AI solution for shipping applications that run entirely on your device.' Uses GPUs and NPUs when available, otherwise the CPU. Aimed at developers building apps.

Price: There are no per-token costs and No Azure subscription is required.

  • Windows
  • macOS
  • Linux

Microsoft Learn · company-reported

Why bother

Why run it yourself?

Privacy, no bill for each message, and working offline, in the makers' own words.

“Ollama runs locally. We don't see your prompts or data when you run locally.”
Ollama docs: FAQ
“Your data never leaves the device, responses start immediately with zero network latency, and your app works offline.”
Microsoft Learn: What is Foundry Local?
“There are no per-token costs and no backend infrastructure to maintain.”
Microsoft Learn: What is Foundry Local?
“Mac Studio lets users run massive models entirely on device with complete privacy”
Apple Newsroom: Apple introduces new Mac Studio with M5 Max and M5 Ultra
“Jan is an open source alternative to ChatGPT that runs 100% offline on your computer.”
Jan on GitHub
“Run large AI models on hardware you own.”
Framework Desktop product page
What we couldn't check

Still to confirm.

Compare all AI models
  • Copilot+ PC minimum RAM (often cited as 16GB) and storageThe Microsoft pages read state only the 40+ TOPS NPU requirement; no RAM or storage minimum was found on them.
  • Apple Mac prices for each memory upgrade (e.g. Mac mini 24GB/32GB, Mac Studio 128GB/256GB/512GB)Apple's store pages load prices with scripts our fetch could not see; only 'starts at' prices from Apple's newsroom are recorded.
  • Current US street prices of RTX 50-series cardsNVIDIA's own pages show placeholder prices; only launch prices from NVIDIA's 2025 newsroom are recorded. Search results mentioned 2026 price rises from memory costs, but no trusted source was read.
  • Apple Intelligence on-device model 2026 update and device listOnly Apple's June 2025 research post was read; a 2026 version could not be checked before the web-fetch budget ran out.
  • DeepSeek V4.1 Flash parameter countsThe model card summary gave conflicting totals (552B backbone vs 763B in metadata); left out of the model list.
  • Kimi K3 License and GLM-5.3 license termsOnly the license names were seen; the license texts were not read.
  • Exact active-parameter counts for Gemma 4 12B and 31B, Qwen 27B modelsModel cards did not state an active count; these are listed as dense / null rather than inferred.
  • Strix Halo machines other than Framework (e.g. HP, ASUS, GMKtec)Not checked before the web-search budget ran out.
  • AMD Variable Graphics Memory (how much of 128GB can be used by the GPU)Not stated on the AMD or Framework pages read; search budget ran out before finding AMD's page on it.
  • MiniMax, Mistral Small 4 / Medium 3.5, Laguna (Poolside) file sizesSeen in listings but not added: LM Studio lists Laguna S 2.1 (118B total, about 8B active) needing 'at least 71 GB of RAM', but the maker card was not read.
  • Independent speed tests on current 2026 models (e.g. Qwen3.8, Gemma 4) on named machinesOnly llama.cpp's own benchmark threads were used; they mostly use the older Llama 2 7B as a fixed test model.