Benchmarks & research
PostHog's Jeeves answers Jev-style decision questions, with an option to work through its reasoning before choosing.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The 9B model adapts Qwen3.5-9B, chooses among allowed answers and drafts reasoning in blocks of four tokens. It uses Jev's question format. PostHog reports better scores than Kev and Jev on its held-out and public JevBench tests, but weaker knowledge results on MMLU. Full reasoning takes much longer. Code is MIT-licensed and weights are Apache-2.0; Kev inspired the project. These are maker tests, not independently reproduced results.
Ask Jeeves which team should handle a customer's message, whether it needs urgent attention and how frustrated the customer is. Its estimates let your program choose among those answers. The README shows running it locally and asking for reasoning too; shorter reasoning limits reduce waiting. Substantial computing memory is needed.