Mulm is a state-of-the-art Hidden Markov Model toolkit, written in Common Lisp. We intend Mulm to be both a solid platform for experimentation and hacking on HMMs in general, and a fast and robust toolkit for PoS tagging. Currently Mulm can decode about as fast as other popular HMM toolkits such as TnT and Hunpos. Decoding speed is mostly dependent on average ambiguity and number of unknown observations. Parameter estimation is still a bit slow. Mulm is tested on LispWorks, Allegro and SBCL and should be reasonably cross-platform. Mulm is primarily developed by Johan Benum Evensberget (johan.benum upon gmail.com) and André Lynum (andrely upon idi.ntnu.no) Mulm includes Mime, a front-end for running experiments and analyzing results. Mime supports regular train and test separation of data and automatic n-way cross-validation. Mulm requires ASDF, CL-PPCRE and Split-Sequence. To load Mulm or Mime for interactive use evaluate (asdf:oos 'asdf:load-op 'mulm) or 'mime at the top level. If you do not have these dependencies we recommend using quicklisp to download them.