Function partial specialization is not yet allowed as per the standard. In the example, you are actually overloading & not specializing the max<T1,T2> function.
Its syntax should have looked somewhat like below, had it been allowed:

// Partial specialization is not allowed by the spec, though!
template <typename T> 
inline T const& max<T,T> (T const& a, T const& b)
{                  ^^^^^ <--- [supposed] specializing here
  return 10;
}

In the case of a function templates, only full specialization is allowed by the C++ standard, -- excluding the compiler extensions!


Since partial specialization is not allowed -- as other answers pointed --, you could work around it using std::is_same and std::enable_if, as below:

template <typename T, class F>
inline typename std::enable_if<std::is_same<T, int>::value, void>::type
typed_foo(const F& f) {
    std::cout << ">>> messing with ints! " << f << std::endl;
}

template <typename T, class F>
inline typename std::enable_if<std::is_same<T, float>::value, void>::type
typed_foo(const F& f) {
    std::cout << ">>> messing with floats! " << f << std::endl;
}

int main(int argc, char *argv[]) {
    typed_foo<int>("works");
    typed_foo<float>(2);
}

Output:

$ ./a.out 
>>> messing with ints! works
>>> messing with floats! 2

Edit: In case you need to be able to treat all the other cases left, you could add a definition which states that already treated cases should not match -- otherwise you'd fall into ambiguous definitions. The definition could be:

template <typename T, class F>
inline typename std::enable_if<(not std::is_same<T, int>::value)
    and (not std::is_same<T, float>::value), void>::type
typed_foo(const F& f) {
    std::cout << ">>> messing with unknown stuff! " << f << std::endl;
}

int main(int argc, char *argv[]) {
    typed_foo<int>("works");
    typed_foo<float>(2);
    typed_foo<std::string>("either");
}

Which produces:

$ ./a.out 
>>> messing with ints! works
>>> messing with floats! 2
>>> messing with unknown stuff! either

Although this all-cases thing looks a bit boring, since you have to tell the compiler everything you've already done, it's quite doable to treat up to 5 or a few more specializations.


What is specialization ?

If you really want to understand templates, you should take a look at functional languages. The world of templates in C++ is a purely functional sublanguage of its own.

In functional languages, selections are done using Pattern Matching:

-- An instance of Maybe is either nothing (None) or something (Just a)
-- where a is any type
data Maybe a = None | Just a

-- declare function isJust, which takes a Maybe
-- and checks whether it's None or Just
isJust :: Maybe a -> Bool

-- definition: two cases (_ is a wildcard)
isJust None = False
isJust Just _ = True

As you can see, we overload the definition of isJust.

Well, C++ class templates work exactly the same way. You provide a main declaration, that states the number and nature of the parameters. It can be just a declaration, or also acts as a definition (your choice), and then you can (if you so wish) provide specializations of the pattern and associate to them a different (otherwise it would be silly) version of the class.

For template functions, specialization is somewhat more awkward: it conflicts somewhat with overload resolution. As such, it has been decided that a specialization would relate to a non-specialized version, and specializations would not be considered during overload resolution. Therefore, the algorithm for selecting the right function becomes:

  1. Perform overload resolution, among regular functions and non-specialized templates
  2. If a non-specialized template is selected, check if a specialization exist for it that would be a better match

(for on in-depth treatment, see GotW #49)

As such, template specialization of functions is a second-zone citizen (literally). As far as I am concerned, we would be better off without them: I have yet to encounter a case where a template specialization use could not be solved with overloading instead.

Is this a template specialization ?

No, it is simply an overload, and this is fine. In fact, overloads usually work as we expect them to, while specializations can be surprising (remember the GotW article I linked).