Free Weights Club: Training Muse Glimmer to out-bet Jev

Fireworks announced its Free Weights Club with free supervised fine-tuning for Muse Glimmer 30B through November 30, presenting an experiment in which the model was trained on resolved Kalshi prediction-market questions to estimate future outcomes without access to market prices or news. Using over 20,000 historical questions for training and separate calibration and evaluation sets, including topic families excluded from training, the tuned model achieved 64.8% accuracy overall and 58.3% on unseen topics, exceeding base Glimmer and Jev by 11.2 percentage points in the latter setting. After calibrating its confidence, tuned Glimmer was correct 86.8% of the time when assigning at least 80% probability, although such high-confidence calls represented a limited share of questions and sometimes involved correlated outcomes. In a hypothetical backtest using historical quotes, modeled fees, and a $1,000 allocation rule, tuned Glimmer returned $1,071 while base Glimmer and Jev lost money, though the company notes that the simulation omitted factors such as liquidity, slippage, reinvestment, and changing fees, and that real-money performance remains uncertain. The post also describes Fireworks’ System One API, which directly scores fixed answer options and achieved faster synthetic-benchmark serving performance than standard chat completions, while directing users to a cookbook and training tools for reproducing the experiment or fine-tuning models for other tasks.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.