SVMPerf - only one vector in the model file? Am I missing something? When I use SVMLight, there's always a lot of vectors in the model file.
But when I create one with SVMPerf - there's only one vector in the file? (?)
But still, it works fine. Am I missing something? 
Also the number of documents that it supposedly used is way smaller than I provided him with ( I supplied 400,000 examples and it says it used 165 training documents? )
 A: The hyperplane for linear SVM can be summarized in a single inner product. This is exactly what libraries like SVMPerf and LIBLINEAR do (which exclusively work with a linear kernel). The resulting hyperplane (e.g. decision function) is exactly the same as the 'original' formulation with a set of support vectors.
The decision function of an SVM model is as follows:
$$f(\mathbf{x}) = \sum_i y_i \alpha_i \kappa(\mathbf{x}_i,\mathbf{x}) + b,$$
where $\mathbf{y}$ is the vector of labels, $\alpha$ the support values, $\mathbf{x}_i$ denotes the i'th support vector, $b$ is a bias term and $\kappa(\cdot,\cdot)$ is the kernel function. Classification is based on the sign of $f(\mathbf{x})$. 
The linear kernel is nothing more than $\kappa(\mathbf{x},\mathbf{y}) = \mathbf{x}^T\mathbf{y}$ assuming you are working with column vectors. Knowing this, it is easy to write the decision function above in terms of a single inner product with the test point:
$$\begin{align}
f(\mathbf{x}) &= \sum_i y_i \alpha_i \mathbf{x}_i^T\mathbf{x} + b, \\
&= \mathbf{w}^T\mathbf{x} + b,
\end{align}$$
with $\mathbf{w} = \sum_i y_i\alpha_i \mathbf{x}_i$.
A: The model computed by SVMLight, is in the dual, i.e., it has to store all support vectors and their weights.
The SVM-Perf works in primal and not dual. That's why only linear kernel is allowed. Thus, it saves a single $w$ as computed by the answer of Marc. 
You see so few documents as Support Vectors because the concepts you try to separate are well defined. For example, I have seen that it takes fewer "football" articles to learn the small concept of "football", than e.g. articles of "politics" to learn the fuzzy concept of "politics". Another reason could be that the size of your documents is a quite large, i.e. you have a dense space.
