Your 5-year-old MacBook can now beat free ChatGPT

A common complaint about running AI on your own computer is that whatever fits is dumber than the free version of ChatGPT. For two years the two were close: the best model a home computer could run and free ChatGPT traded places by a few points at most. Now it isn’t close.

In August, Alibaba released Qwen3.8 27B, an open model small enough to run on a five-year-old MacBook Pro or a gaming PC with a 24GB graphics card. It scores 33.7 on the Artificial Analysis Intelligence Index. Free ChatGPT, which switched to GPT-6 Luna on 8 October, scores 18.5, or 21.5 when you press Think.

The lead has changed hands

Since September 2024 the lead has changed hands 7 times. In February 2026, Alibaba’s Qwen3.5 27B scored 22.9 to the 17 of free ChatGPT’s GPT-5.2, the widest gap until then. In June, OpenAI updated GPT-5.5 Instant, free users’ default since May; the new version scored 26 and took the lead back.

Then OpenAI changed what free means. On 6 August it made GPT-5.6 Luna the default for free users, with no cap on text chats and a new Think button. Its default answer scores 15.5, lower than the model it replaced; with Think, 21. That put the Qwen3.5 27B already on the Mac back in front the same day. Eight days later, Qwen3.8 27B widened the lead to 33.7.

Qwen3.8 27B vs GPT-6 Luna: read the matchup

What your computer can run

What fits depends on memory. A graphics card can give a model all of its own memory; a Mac shares its memory with the rest of the computer, and macOS lets the graphics chip use about two thirds of it.

On a 16GB MacBook, or an 8GB or 16GB graphics card, the best option is K2 Horizon 7B, released in September by Institute of Foundation Models in Abu Dhabi. It needs about 6GB and scores 20.6: ahead of free ChatGPT’s default answer, a point behind it with Think. A 24GB card or a 32GB MacBook Pro runs Qwen3.8 27B, which needs about 18GB.

More memory doesn’t help, until it’s a lot more

A 32GB graphics card, the RTX 5090, runs the same Qwen3.8 27B as a 24GB one. Labs make open models either around 30 billion parameters, to fit a gaming card, or well past 100 billion, for data centers. Of the open models released in the past year, the smallest that beats Qwen3.8 27B is Qwen3.8-Flash-Next, which needs about 110GB.

How far behind the best

Beating free ChatGPT is not the same as matching the best. The highest score today belongs to Claude Opus 5.5, at 57.6. The best open model of any size, MiMo-V2.6-Pro, scores 46.3 and needs a server.

A 24GB card or a 32GB MacBook Pro runs what the best model anywhere could do in March 2026, when GPT-5.4 came out: about 7 months behind. A 16GB machine is about 17 months behind, at the level of Claude 4 Opus.

See the top AI model on any date

The catch

Qwen3.8 27B’s score comes from thinking at length before it answers, and a home computer thinks slowly. Each word it writes means reading the whole model from memory, so speed comes down to how fast the machine’s memory is.

By our estimate, a 2021 MacBook Pro with an M1 Pro runs Qwen3.8 27B at about 8 tokens a second, so a hard question that takes 2,000 tokens of thinking and answer takes about 4 minutes. The M1 Max, with twice the memory speed, does it in about 2 minutes. An RTX 4090 manages about 40 tokens a second; free ChatGPT’s model streamed about 97 in our tests. K2 Horizon 7B is quick almost everywhere. With thinking turned off, Qwen3.8 27B scores 20.2: still ahead of free ChatGPT’s default, behind it with Think.

Memory also goes to the conversation. Our numbers leave room for about 8,000 tokens of it; a long document or chat needs several more gigabytes, which on a 24GB card means a smaller version of the model. And the scores come from the full models, while a home computer runs a compressed copy that is usually a little worse.

When will it be as fast?

That 6 times is for the smartest model the Mac can run. If you only need free ChatGPT’s level, the gap is already smaller, and it keeps shrinking, because each new batch of small models fits the same intelligence into fewer parameters, and fewer parameters run faster.

GPT-5 became free ChatGPT in August 2025. By September 2025, a 2021 MacBook Pro could match it with Apriel-v1.5-15B-Thinker, at about 29 tokens a second; 5 months later Qwen3.5 4B matched it at 108, faster than free ChatGPT runs today. GPT-5.1’s level took 11 months from the first model that matched it. Today’s level, GPT-6 Luna’s 18.5, first fit the Mac in February 2026 with Qwen3.5 27B at 16 tokens a second. Since September 2026, K2 Horizon 7B matches it at 62.

If today’s level takes as long as GPT-5.1’s did, a model as smart as free ChatGPT will run as fast on a five-year-old MacBook by about January 2027. Faster chips would only bring that closer. Free ChatGPT may get smarter or faster in the meantime, so the target moves too.

What a few hundred dollars buys

If you don’t already own the computer, the cheapest way past free ChatGPT, even with Think, is a used graphics card. A used RTX 3090 24GB, about $800, runs Qwen3.8 27B. The next step up in intelligence costs about $3,999, for an NVIDIA DGX Spark with 128GB that runs Qwen3.8-Flash-Next (39.8 points against 33.7).

Short of that, spending more buys the same model at best: a $1,999 RTX 5090 runs Qwen3.8 27B faster, not smarter, and a $3,299 mini PC with 128GB runs it slower, because its memory is slower. At the low end, a $429 RTX 5060 Ti or a $899 Mac mini with 16GB runs K2 Horizon 7B, ahead of free ChatGPT’s default answer but not of Think. And no price on this chart reaches Claude Opus 5.5, which scores 57.6.

None of that changes the direction. A year ago, the best model a home computer could run and free ChatGPT were within a few points of each other. Today the home model leads by 12 points even when you press Think, costs nothing per question, and never sends your words anywhere.