I think it's important to keep two things separate here:

First, if $(M,F)$ is a smooth Finsler manifold (which means that $F^2:TM\to [0,\infty)$ is smooth and strongly convex away from the zero section of $TM$), then the unit sphere bundle $\Sigma\subset TM$ (aka the *tangent indicatrix*) is a smooth hypersurface in $TM$ and there is an open neighborhood $U$ of $\Sigma\times\{0\}$ in $\Sigma\times\mathbb{R}$ on which there is defined a smooth mapping $\exp:U\to M$ such that, for each fixed $u\in \Sigma$, the curve $\gamma_u(t) = \exp(u,t)$ (defined for all $t\in\mathbb{R}$ such that $(u,t)$ lies in $U$) is the maximally extended unit speed $F$-geodesic with initial velocity $u\in\Sigma$.

Second, for any given $p\in M$, there is a $\delta>0$ such that, if $B_\delta(p)\subset T_pM$ is the set of vectors $v\in T_pM$ satisfying $F(v)<\delta$, then there is a well-defined mapping $\exp_p:B_\delta(p)\to M$ such that $\exp_p(0_p) = p$ and $\exp_p(tu) = \exp(u,t)$ for all $u\in \Sigma_p$ and all $t\in(0,\delta)$. By taking $\delta$ sufficiently small, one can ensure that $\exp_p:B_\delta(p)\to M$ is a homeomorphism onto its (open) image, one that is a smooth diffeomorphism away from $0_p$.

Now, unless $F$ is reversible (i.e., $F(-v) = F(v)$), the map $\exp_p:B_\delta(p)\to M$ need not be $C^2$ at $0_p$, because it can easily happen that $\exp(-u,-t)\not=\exp(u,t)$, so the $\exp_p$-image of the line segment $\{tu\ |\ |t|<\delta\}$ need not be a $C^2$ curve in $M$ (it will be $C^1$, though).
Even when $F$ is reversible (or, more generally, geodesically reversible), so that the exponential map $\exp_p$ is smooth on lines through $0_p\in B_\delta(p)$, it may still not be smooth at $0_p$.

Unfortunately, writing down a simple, explicit example for which smoothness clearly fails is not easy because, usually, it is not possible to integrate the geodesic equations and compute the map $\exp:U\to M$ explicitly. In one case where it is possible, namely the case of a Minkowski Finsler metric (i.e., $M=\mathbb{R}^n$ and $F$ is invariant under translations in space), the exponential map
$$\exp:\Sigma\times \mathbb{R} \bigl(= (\mathbb{R}^n\times\Sigma_0)\times\mathbb{R}\bigr)\to\mathbb{R}^n$$
is $\exp((p,u),t) = p+tu$ for $p\in \mathbb{R}^n$ and $u\in \Sigma_0\subset T_0\mathbb{R}^n = \mathbb{R}^n$, so the map $\exp_p:T_p\mathbb{R}^n\to\mathbb{R}^n$ is a smooth diffeomorphism after all, regardless of how non-Riemannian $F$ is. (This shows that it's not just a matter of looking at the shapes of the unit spheres of $F$. The failure of smoothness is more subtle than that.)

However, in an answer to this question, I gave an example of a (homogeneous!) Finsler metric $F$ on a surface $M$ that is reversible and has the property that $F^4$ is a smooth function on $TM$, but the fourth power of the Finsler distance function from any given point of $M$ is *not* smooth. Thus, this is an example in which the map $\exp_p$ cannot be smooth at the origin $0_p$ (for any $p\in M$), because, if $s_p:M\to[0,\infty)$ is the function that gives the $F$-distance from $p\in M$, then, for $\delta>0$ sufficiently small, we have $F(v) = s_p\bigl(\exp_p(v)\bigr)$ for all $v\in B_\delta(p)\subset T_pM$.

**Remark (28 Feb 2016):** I was curious as to what are the necessary and sufficient conditions for a Finsler surface to have $\exp_p$ be at least $C^2$ at every point. It turns out that this is very restrictive: Either the Finsler structure must be Minkowskian (i.e., invariant under a $2$-dimensional abelian group of translations as discussed above) or else the Cartan scalar (usually denoted $I$) must be constant. The latter case includes the Riemannian case ($I$=0) as a special case; the cases where $I$ is a non-zero constant are actually 'generalized' Finsler structures (at each $p\in M$, the unit tangent vectors in $T_pM$ form a logarithmic spiral rather than a closed convex curve). In all these cases, the map $\exp_p$ is actually $C^\infty$.