Can a 14GB file on a mid-range gaming laptop out-recommend the algorithm of the world's largest PC games store? According to XDA Developers writer Abhinav Raj, it did, at least for him. In his single-user test, a local Gemma model suggested more games he would play than Steam's own Interactive Recommender. The result is entertaining and a little provocative. It is also an anecdote, and the details matter more than the headline.
The setup: 25 games, one CSV, no internet
Gemma 4 comes in five sizes: E2B, E4B, 12B, 26B A4B and 31B. Raj used the 26B A4B Instruct variant. It is a Mixture-of-Experts model with 25.2B total parameters and 3.8B active parameters. Because only a 4B subset of parameters is active during inference, Google says it runs much faster than its 26B total would suggest. That helps explain why a model this size is practical on consumer hardware.
Raj's reported configuration:
- Model file: bartowski's IQ4_XS GGUF quantization, 14.21GB.
- Runtime: llama.cpp.
- Sampler: temperature 1.0, top_p 0.95, top_k 64.
- Reasoning: enabled.
- Web search: none, so the model worked from training data alone.
- Hardware: an older RTX 3070, according to his closing remarks.
- Input: a CSV exported from Steam with 25 owned titles, playtime hours and last-played dates. It had no other personal information.
- Task: recommend exactly 28 games he doesn't own, to match the length of Steam's list. Each pick needed its exact Steam name and a one-sentence justification that cited specific games in his CSV.
Raj says the model first sorted his library into playtime tiers, then into five genre buckets, and only then looked for matches.
Two details are missing from the account. It doesn't report memory use, context settings, or inference speed. It also doesn't say how the work was split between GPU and CPU. A 14.21GB file is not the same as the memory a full runtime needs, so readers shouldn't treat it as a minimum hardware spec.
What Steam's recommender actually does
Valve introduced the Interactive Recommender on March 18, 2020, after it started as Steam Labs Experiment 002. Valve's announcement says it uses a machine learning model trained on the playtime histories of millions of Steam users. Valve also says the system is not directly affected by tags or reviews. The model learns about games from what people actually play. The premise is that if players with habits like yours enjoy a game you haven't tried, you probably will too.
Valve is also open about the limits. The recommender can't surface new releases that nobody has played yet, and the Discovery Queue handles that job. Valve says the same underlying model also powers "Play Next," which recommends games you already own but haven't played.
Valve's figures on how well it worked are engagement numbers. It measured clickthrough and wishlist and purchase conversion, and said more than 10,000 different games were bought after visits to the recommender page. Those figures say nothing about how it compares with a language model.
This design explains the mismatch. Steam's system uses aggregated behavior from millions of players. Gemma saw one person's 25 lines and wrote a rationale for each pick. One is a population model, and the other is closer to a well-read friend skimming your shelf.
The results, and how to read them
Raj counted the recommendations he would happily play:
| System | Picks he'd play | Share of 28 |
|---|---|---|
| Gemma 4 26B A4B (local) | 16 | about 57% |
| Steam Interactive Recommender | 11 | about 39% |
That is a five-game gap, and the rating is entirely his own. There was no blinded scoring, no second rater, and no repeat run. His library is also small. It is not a benchmark, and it doesn't show that local models beat Steam in general.
Some specifics from his write-up:
- Where Steam did well: R.E.P.O. and Warframe landed. Hogwarts Legacy and Red Dead Redemption 2 were described as spot-on, even though Raj owns both on the Epic Games Store.
- Where Steam missed: Yu-Gi-Oh! Master Duel, which ranked in the top five, and Dota 2. Raj saw nothing in his library to suggest either.
- Where Gemma did well: the racing picks. Assetto Corsa Competizione and F1 24 have both been on his wishlist for months. He guesses the model inferred them from his NASCAR and Trackmania play.
- Where Gemma missed: Hades looked like a stretch for someone playing Digimon Story: Time Stranger. Half-Life: Alyx made little sense for a player without a VR headset, but the model was never told that. It also recommended Forza Horizon 5, which he already owns, even though the prompt asked only for unowned games.
The Forza slip matters. LLMs don't reliably follow hard constraints, even when the CSV makes the constraint checkable. Anyone repeating this should check each suggestion against their library. They should also confirm that the exact game names exist on Steam, because a model working from memory can misname or invent titles.
Missing data is a bigger problem than model quality
The most useful part of the story may be what neither system could see. Raj has about 200 hours in Microsoft Flight Simulator 2024 through Game Pass. Steam can't see that, and neither could Gemma, because he fed it only his Steam export. His Steam library holds Flight Simulator X, so the recommender only roughly picked up on his interest in flight simulators and offered War Thunder. He would have preferred DCS World.
That points to a practical lesson for PC gamers who split libraries across Steam, Epic and Game Pass:
- A recommender is only as good as the history it sees.
- A local model lets you add what the store doesn't know, such as Game Pass hours, Epic purchases, hardware limits, or genres you're sick of.
- Adding that context would probably have fixed several of the misses above. In particular, a line saying "no VR headset" would have been enough to avoid Half-Life: Alyx. That is my inference from the failure, not something Raj tested.
What this does and doesn't prove
- Fair: a small local model can write personalized, explained recommendations from a short playtime CSV. In this test, more of them suited the author than Steam's list did.
- Not established: that it generalizes. Other users, bigger libraries or different runs could land differently, and the sampler uses temperature 1.0, so outputs will vary between runs.
- Not a like-for-like contest: the two systems have different inputs and different goals. Steam can't recommend unreleased or unplayed titles by design, and it doesn't take free-text instructions.
- Privacy: running locally with no web search means the library data stays on that setup, according to the author's description. That is a reasonable expectation, but it isn't a blanket guarantee about every piece of software in the stack.
Raj's closing argument is that this is the worst local models will be at the task. Nothing in his data tests that claim, but the underlying trend is plausible. The model he used, a 25.2B-parameter MoE with 3.8B active parameters, is now small enough to run on a laptop with a three-generation-old GPU.
Bottom line
This is a fun, low-stakes experiment, and a useful reminder of two things. Personalization depends on what a system is allowed to know, and 11 of 28 versus 16 of 28 is a taste test, not a benchmark. If you want to try it yourself, export your library with playtime, tell the model what the store can't see, and verify every pick. Treat the output as a prompt for your wishlist, not a verdict on Valve's algorithm.