ExpectedHarm = P_incident * Harm
Weapons multiply harm, and education reduces P. Nuclear weapons cause ExpectedHarm to approach infinity almost regardless of P_incident, so they are out. But guns, knives, acid, and other weapons all sit along a continuum.
Increasing access to guns unquestionably increases ExpectedHarm (arguments like 'if everyone had guns, people would be well behaved' are, of course, ridiculous on their face), but there are social benefits to guns. Some people enjoy hunting, sport shooting, and carrying a gun makes many people 'feel' safer (which, though hard to quantify, remains a true social good, in some sense). Guns are also, to some degree, a check on government power. I don't particularly like the Bundy's position or what they stood for, but their access to guns is absolutely what gave them the power to make that stand. You can count this as a positive or negative depending on your perspective. I tend to count it as a positive even though most uses of it, like the Bundy's, I strongly disagree with.
And finally, there is the sort of 'constant' boost to the positive side you get from the general bias we ought to have against banning things (you may choose to set this to zero or negative if you hate freedom). So, the full equation is:
SocialGood + LibertarianBias - P_incident * Harm * BanEfficacy
If this expression is greater than zero, you should leave it alone. This does ignore some theoretical nonlinear terms (e.g. banning guns increases the probability of other forms of violence, or vice versa), but it more or less captures the idea. Depending on which values you input for various terms, you'll get a different answer.
All of that is to say that whether or not any given weapon ought to be legal should be evaluated according to a framework that looks something like the above.