I gave my local LLM my entire Steam library, and it recommends games better than Steam does

Abhinav pivoted from a career in banking to pursue his first love in writing. Even while working full-time, he continued contributing as an editor-at-large, a role he has held for more than 7 years. A lifelong tech enthusiast who has built three gaming and productivity powerhouse PCs since 2018, his passion for technology keeps him closely following the semiconductor industry, from NVIDIA and AMD to ARM. His MSc dissertation explored how artificial intelligence will reshape the future of work, reflecting his curiosity about the wider social impact of emerging technologies.

Local LLMs have become faster, smarter, and more reliable than perhaps ever in human history. I've previously talked about how some capable 27B models have gone from merely complementing my coding workflow to replacing some of the paid tools I had been using, and the local AI landscape is moving so fast that I'd wager the models have only gotten better at what they do since then.

If such models can boost my productivity enough in a way that warrants me writing entire essays about them, recommending a few games that I'd enjoy based on what I play shouldn't be much of a stretch, right?

Steam already does it, but the problem is that its recommendations are based on how long I played something, and how likely another user with similar habits is to play a game that I haven't tried. It's depersonalized by design, and that's why I decided to see if a local model could do a better job at it.

How intelligent is Steam's Interactive Recommender?

It knows players like me, but it isn't the same as knowing me

Steam's Interactive Recommender was introduced in 2020 as a part of the Steam Labs initiative, and Valve put it forward as a way for users to discover personalized recommendations that draw from their playing patterns. As expected, it uses a machine learning algorithm for these recommendations that is trained "based on the playtime histories of millions of Steam users."

At a cursory glance, it does exactly what I expected of it. Several dozen recommendations were listed right next to the titles that I engaged with most frequently. A few of the picks actually landed quite well. For example, R.E.P.O. is a game that I've been pushing my friends to play for weeks, and Warframe is a title that I grew up on when I just had a laptop with Intel HD Graphics 4000.

Hogwarts Legacy and Red Dead Redemption 2 were both spot-on. I already own both on the Epic Games Store, and I am currently on my second playthrough of Hogwarts Legacy. That's highly impressive considering that Steam managed to get these right without knowing I'd played them.

The rest, however, is a bit of a mess. Of the 28 recommendations, most are games that I'd probably never touch. My Steam library has Flight Simulator X, but the 200-odd hours I've spent in Microsoft Flight Simulator 2024 are on Game Pass, which is invisible to Steam. The closest it gets to recognizing my flight sim obsession is War Thunder, which technically is a simulation title but is combat-first. DCS World would've been a closer match, especially since it's right there on the store.

Then, there are titles like Yu-Gi-Oh! Master Duel and Dota 2, neither of which I've shown any interest in playing, and I'd argue that there's nothing in my library that suggests I'd enjoy those either. The former was in the top 5 of recommendations, which makes me trust the algorithm a lot less.

I handed my entire library to Gemma 4 26B A4B

And hoped for better recommendations to follow

For the purpose of this test, I used Google's Gemma 4 26B A4B Instruct in bartowski's IQ4_XS GGUF, which was a 14.21GB file served through llama.cpp harness, along with Google's standard sampler values of temperature 1.0, top_p 0.95, and top_k 64. I took the same CSV that I fetched from Steam with 25 owned titles, playtime hours, and no other additional information. To make the comparison even, I asked it to recommend 28 titles to match Steam's list.

The prompt was straightforward and simple:

You are a game recommender. You will be given my Steam library as a CSV with playtime hours and last played date. Recommend games from outside my library that fit my taste. Recommend exactly 28 games that I don't own. For each pick, give the game's exact name as it appears on Steam, followed by one sentence explaining the pick with reference to specific games in my CSV.

After delivering the prompt, I waited and watched its reasoning, as it was set as enabled. There was no web search, so it was working off its training data alone. It started by sorting my library into tiers of playtime, then into five genre buckets, and only then went looking for picks to fit each one.

The local model outperformed Steam's recommendation system

Out of the 28 recommended, I would happily play 16

Gemma 4 got 16 out of 28 recommendations right, which gives it a success rate of 57% compared to Steam's 11 out of 28, or 39%. That's a huge margin of victory for a 14GB model running entirely on my laptop with nothing more than the 25 lines in the CSV that I gave it to work with. Gemma 4 managed to recommend almost twice as many games I'd consider playing, which is about five more than the store I buy those games from.

The best of the bunch were the racing simulation recommendations. Forza Horizon 5 was, again, a title I already owned, whereas Assetto Corsa Competizione, and F1 24 have both been on my wishlist for months. This was likely picked up on the basis of my interest in NASCAR and Trackmania, but then again, that's also information that Steam also happens to have.

This is not to say that the model was perfect, though. The model's recommendation of Hades to someone who was interested in Digimon Story: Time Stranger was a bit of a stretch, especially considering they're not even in the same genre. Half-Life: Alyx was also slightly out of left field for someone without a VR headset, but hey, the model was never told that.

Gemma 4 was better, and this is the worst it'll ever be

Perhaps the one aspect where the model impressed me the most, was that, Gemma 4's recommendations felt more personal rather than general, and that's something I hypothesized before the start of the test. If local models on an old RTX 3070 are already this close on a task that Steam probably built an entire ML team around, the gap is only going to close from here, and that's something that makes me genuinely excited for the future of local AI inference.