This study investigates affective polarization among Swedish politicians on Twitter from 2021 to 2023, including the September 2022 parliamentary election. Analyzing over 25,000 tweets and employing large language models (LLMs) for sentiment and political classification, we distinguish between positive partisanship (support of allies) and negative partisanship (criticism of opponents). Our findings are contingent on the definition of the in-group. When political in-groups are defined at the ideological bloc level, negative and positive partisanship occur at similar rates. However, when the in-group is defined at the party level, negative partisanship becomes significantly more dominant and is 1.51 times more likely (1.45, 1.58). This effect is even stronger among extreme politicians, who engage in negativity more than their moderate counterparts. Negative partisanship also proves to be a strategic choice for online visibility, attracting 3.18 more likes and 1.69 more retweets on average. By adapting methods developed for two-party systems and leveraging LLMs for Swedish-language analysis, we provide novel insights into how multiparty politics shapes polarizing discourse. Our results underscore both the strategic appeal of negativity in digital spaces and the growing potential of LLMs for large-scale, non-English political research.