One player said he’d play 16 Gemma picks and 11 from Steam

Abhinav Raj said he would play 16 of Gemma’s 28 suggestions, compared with 11 of Steam’s 28. The result comes from one person’s judgment in a comparison published by XDA Developers on October 13, 2026—not a general measure of how well local AI recommends games. XDA Developers’ account

What the numbers say

Raj assessed 28 suggestions from each system using the same personal standard: whether he would play the game. His counts work out to 57.1% for Gemma and 39.3% for Steam, a difference of 17.9 percentage points. Those percentages describe Raj’s picks in this comparison, not a score that predicts what another player will like.

The examples cut both ways. Raj considered Assetto Corsa Competizione and F1 24 good fits, and said they were already on his wishlist. He also questioned Hades as a match for Digimon Story: Time Stranger and noted that Half-Life: Alyx would require a VR headset. Steam’s list included recommendations he liked, such as R.E.P.O., Warframe, Hogwarts Legacy and Red Dead Redemption 2, alongside picks he considered poor fits, including Yu-Gi-Oh! Master Duel and Dota 2. These are examples from Raj’s experience, not a representative sample of players.

What Gemma received

Raj used a CSV with 25 owned Steam games and their playtime hours as context. He asked Gemma for exactly 28 games outside that library, with an explanation connecting each suggestion to games in the CSV; he said web search was off and that he gave the model no other additional information.

The local setup used Google’s Gemma 4 26B A4B Instruct, served through llama.cpp with a bartowski IQ4_XS GGUF model file reported as 14.21 GB. That is the file’s size. The XDA account describes one person’s recommendation run, so the outcome does not establish that Gemma will suit another player’s library or preferences.

What Steam’s recommender is designed to do

Valve describes the Steam Interactive Recommender as a machine-learning system based on the playtime histories of millions of Steam users. Valve says tags and reviews do not directly affect this recommender. Its filters let users balance popular against niche games and recent releases against classics. Valve’s description of the Interactive Recommender

That feature is one part of Steam’s discovery tools. Valve says the Interactive Recommender cannot surface new releases that nobody has played; the Discovery Queue is the feature it points to for finding new releases. Raj also noted that Steam could not account for his Microsoft Flight Simulator 2024 playtime through Game Pass or games he owned through the Epic Games Store—context specific to his comparison.