How small can a smart model get?

Open-weight model size vs. capability, Dec 2024 – Jul 2026

With the release of DeepSeek V4 I wanted to see how model sizes are trending over time. What surprised me is that DeepSeek V4 Flash is extremely capable and small enough to run on a setup you could build for under $50,000 — expensive, but not a datacentre. So I pulled together the open-weight models with published parameter counts and their current Artificial Analysis Intelligence Index scores, and made the two plots below with Opus/Sonnet 5.

I am not an expert in LLMs, and there are surely physical limits to how much you can squeeze into a small model. But I don't know how close we are to them — small models may have a long way to go before saturating. If the trend holds, we should have Opus 4.5–level models (roughly an AA score near 40) running on an ordinary MacBook Air or Pro by this time next year.

AA score over time by parameter-count bin — **Figure 1.** Score against release date for 101 open-weight models, grouped by total parameter count; marker area is the actual size. The orange step is the running maximum over closed frontier models, which don't publish their sizes. Every bin is climbing, and the small ones are climbing fastest.

Iso-score contours over time and model size — **Figure 2.** The same data as a fitted surface: contours of equal score over time and model size. They fall about 1.7 decades per year — roughly 54× fewer parameters for the same score, one year later. Black steps are the smallest model actually observed at each score, with no fitting; dashed lines are what fits in a given device at 4-bit.

Caveats: I don't trust the trend above a score of 40 — there are very few data points up there. Some scores are Artificial Analysis estimates rather than measured runs, reasoning and non-reasoning variants are a several-point confound, and release date is not the same as training-compute date.