How long would it take to crack passwords up to 10 letters long if you knew the hashes were MD5 with a salt (that you know)?
Computation is simple enough, though there are some hidden assumptions.
On this site, one can find benchmarks for some GPU-based cracking systems. One of them, featuring no less then eight AMD R9 290X GPU, can compute 93.8 billions of MD5 per second.
There are 2610 = 141167095653376 possible sequences of 10 letters. At 93.8 billions per second, they can all be hashed in about 1500 seconds, i.e. 25 minutes. On average, finding the right password will take half that time (sometimes you get lucky, sometimes you don't).
However, mind the details:
MD5 is a hash function. Its standard definition does not speak of passwords or salts. When you have "passwords hashed with MD5 and a salt", then you really are using some unspecified algorithm which uses MD5 as one of its internal elements. Whether the computational cost of the algorithm can be reduced to "just one MD5" depends on that algorithm. Some MD5-based password hashing functions actually imply several MD5 invocations, possibly a lot (in that sense, PBKDF2 with HMAC/MD5 is an algorithm "based on MD5" and can involve billions of invocations for each password).
The computation above assumes 10 lowercase letters. If the password may include both lowercase and uppercase letters, and they are treated differently, then the number of combinations rises to 5210 = 144555105949057024, for a corresponding brute force time of 17.8 days. If we use printable characters (the 95 ASCII signs, excluding control characters but including space), the number of combinations rises again, up to 9510 = 59873693923837890625, for a brute force time of about 20 years (again, this is for the complete space exploration; attack time will be half that value on average).
Conversely, when normal human users are supposed to generate a password, they come up with "witty" passwords, not "random" passwords. Among 10-letter possible passwords, some are widely more probable than others. Attackers will try the passwords by beginning with the more probable ones, and this reduces attack time quite a lot.
The system used here as example is quite big in amateur terms (8 big GPU...), but an industrious attacker may accumulate more power. He may also rent it (from commercial cloud systems), which can be cost effective. There is no absolute answer to your question if you do not first define the attacker's budget.