diff --git a/runner/ollamarunner/runner.go b/runner/ollamarunner/runner.go
index f3286abae..0a14e0073 100644
--- a/runner/ollamarunner/runner.go
+++ b/runner/ollamarunner/runner.go
@@ -298,12 +298,6 @@ type Server struct {
 	// multimodalHash generates hashes for comparing equality
 	// of non-text data
 	multimodalHash maphash.Hash
-
-	// vocab is a llama.cpp vocab required for gammar-based
-	// constrained generation (json mode, structured outputs)
-	// TODO: this is temporary until Ollama sampling supports
-	// constrained generation
-	vocab *sample.Vocab
 }
 
 func (s *Server) allNil() bool {
@@ -609,7 +603,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	var grammar *sample.Grammar
 	var err error
 	if req.Grammar != "" {
-		grammar, err = sample.NewGrammar(s.vocab, req.Grammar)
+		grammar, err = sample.NewGrammar(s.model.(model.TextProcessor).Vocabulary(), req.Grammar)
 		if err != nil {
 			http.Error(w, "failed to load model vocabulary required for format", http.StatusInternalServerError)
 			return
@@ -749,8 +743,6 @@ func (s *Server) loadModel(
 		panic(err)
 	}
 
-	s.vocab = sample.NewVocab(mpath)
-
 	// TODO(jessegross): LoRA loading
 	if lpath.String() != "" {
 		panic("loras are not yet implemented")