ModelCompare
Compare

Open-source AI models

What's the best AI model your laptop can run?

On a 16GB Mac, the smartest open model you can run is K2 Horizon 7B. It writes about 19 tokens a second, faster than you read, and it's smarter than free ChatGPT.

Try GPT OSS 20B at your speed See all 16GB Mac models
The smartest open model each machine runs
  1. 15.6K2 Horizon 3.7B18 tok/s
  2. 20.6K2 Horizon 7B12–69 tok/s
  3. 33.7Qwen3.8 27B10–72 tok/s
  4. 39.8Qwen3.8-Flash-Next81 tok/s
  5. 43.9Ling 3.1 Flash20 tok/s
  1. 57.6Claude Opus 5.584 tok/s via APIBest of any kind, closed
  2. 18.5GPT-6 Luna97 tok/s via APIFree ChatGPT, closed

Artificial Analysis Intelligence Index. Speed is our estimate at 4-bit from the machine's memory bandwidth, not a measurement. For scale, Claude Opus 5.5 writes about 84 tokens a second through an API.

Finder

Pick a job and a machine

Scores are Artificial Analysis evaluations (0 to 100). Memory is at 4-bit with an 8k context. Prices are per million tokens, 3 parts input to 1 part output.

What is it for?
Where will it run?

The best open model overall is Kimi K3.1 Pro from Kimi, scoring 55.1. 203 models have a score.

For scale, through an API on ModelCompare: Claude Opus 5.5 84 tokens/s · GPT-6 Luna (free ChatGPT) 97 tokens/s · Claude Sonnet 5.5 100 tokens/s · Gemini 3.8 Flash 183 tokens/s. Reading speed is about 10.

  1. 1
    Kimi K3.1 ProKimi · Oct 2026
    Overall score55.1
    Coding–
    Agents–
    Science–
    Reasoning52.5
    Memory
    Unknown
    size not published
    Through an API
    $4/M
    111 tok/s
    Try it now Weights
  2. 2
    MiMo V2.6 ProXiaomi · Sep 2026 · 1.0T (42B active) · MIT
    Overall score46.3
    Coding–
    Agents–
    Science–
    Reasoning49.4
    Memory
    614 GB
    server-class
    Through an API
    $0.54/M
    27 tok/s
    Try it now Weights
  3. 3
    DeepSeek V4.1 FlashDeepSeek · Oct 2026 · 763B (16B active) · MIT
    Overall score45.1
    Coding–
    Agents–
    Science–
    Reasoning43.3
    Memory
    459 GB
    server-class
    Through an API
    $0.26/M
    155 tok/s
    Try it now Weights
  4. 4
    GLM 5.3Z.ai · Aug 2026 · 744B (40B active) · glm-5.3
    Overall score44.8
    Coding74.8
    Agents53.1
    Science91.7
    Reasoning42.3
    Memory
    448 GB
    server-class
    Through an API
    $1.90/M
    111 tok/s
    Try it now Weights
  5. 5
    Ling 3.1 FlashInclusionAI · Sep 2026 · 560B (25B active) · MIT
    Overall score43.9
    Coding–
    Agents–
    Science–
    Reasoning43.5
    Memory
    338 GB
    512GB Mac and up
    Through an API
    $0.45/M
    Try it now WeightsOllama

How much smarter open models got

A year ago the best open model was GLM-4.6 at 18.5. Today it's MiMo-V2.6-Pro at 46.3, 2.5 times the score.

0306020252026
Today: MiMo-V2.6-Pro at 46.3, against 57.6 for Claude Opus 5.5 and 18.5 for GPT-6 Luna (free ChatGPT). Artificial Analysis Intelligence Index with today's tests, by release date.

Cost

Running it yourself vs paying per token

For each machine, the best model it runs that you can also rent through an API. Running it yourself costs electricity (at $0.17 per kWh) plus the machine. The API costs per token but answers far faster.

MachineBest model with an APISpeedElectricity /MAPI /MMachine pays for itself after
8GB cardabout $299Qwen3.5 9B33 tok/s$0.25$0.19Never: the API is cheaper
16GB Macabout $899Gemma 4 12B11 tok/s$0.13$0.1542 billion tokens122 years nonstop
16GB cardabout $429Gemma 4 12B40 tok/s$0.28$0.15Never: the API is cheaper
32GB Macabout $3,199Qwen3.8 27B16 tok/s$0.24$1.133.6 billion tokens7 years nonstop
24GB cardabout $800Qwen3.8 27B38 tok/s$0.51$1.131.3 billion tokens396 days nonstop
32GB cardabout $1,999Qwen3.8 27B72 tok/s$0.42$1.132.8 billion tokens454 days nonstop
64GB Macabout $2,699Qwen3.8 27B22 tok/s$0.21$1.133.0 billion tokens4 years nonstop
128GB mini PCabout $3,999Qwen3.8 27B10 tok/s$0.94$1.1322 billion tokens70 years nonstop
128GB Macabout $4,699Qwen3.8 27B22 tok/s$0.28$1.135.6 billion tokens8 years nonstop
256GB Macabout $5,500Qwen3.8-Flash-Next81 tok/s$0.13$0.2354 billion tokens21 years nonstop
512GB Macabout $9,499Ling 3.1 Flash20 tok/s$0.64$0.45Never: the API is cheaper

Prices are rough US prices; a graphics card also needs a PC around it. Power is the whole machine while it writes. Electricity counts written tokens only (reading your prompt is much faster), while the API price mixes 3 parts input to 1 part output. Big mixture-of-experts models whose active size isn't published here are left out, since their speed can't be estimated. Running it yourself also means your prompts never leave your machine, and there's no rate limit.

Run one yourself

Four ways to run an open model yourself

Ollama

The quickest start. Install it, then type ollama run and a model name in a terminal. It downloads a 4-bit build and chats right away, and serves an API other apps can use.

LM Studio

A desktop app with no terminal: search for a model, download it and chat. It shows which builds fit your memory and uses MLX on Macs.

On a phone

PocketPal AI (iPhone and Android) downloads small GGUF models from Hugging Face and chats offline. Google's AI Edge Gallery runs Gemma models on Android and iPhone.

llama.cpp, MLX and vLLM

For more control. llama.cpp runs GGUF files anywhere, MLX is Apple's fast path on Macs, and vLLM serves many users at once from NVIDIA GPUs.

How far a laptop has come: Your 5-year-old MacBook can now beat free ChatGPT.