diff --git a/fs/config.go b/fs/config.go
new file mode 100644
index 000000000..bc5bfa550
--- /dev/null
+++ b/fs/config.go
@@ -0,0 +1,13 @@
+package fs
+
+type Config interface {
+	Architecture() string
+	String(string, ...string) string
+	Uint(string, ...uint32) uint32
+	Float(string, ...float32) float32
+	Bool(string, ...bool) bool
+
+	Strings(string, ...[]string) []string
+	Uints(string, ...[]uint32) []uint32
+	Floats(string, ...[]float32) []float32
+}
diff --git a/kvcache/causal_test.go b/kvcache/causal_test.go
index bf98abef6..517e3726d 100644
--- a/kvcache/causal_test.go
+++ b/kvcache/causal_test.go
@@ -5,6 +5,7 @@ import (
 	"slices"
 	"testing"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model/input"
 )
@@ -373,7 +374,7 @@ func TestCanResume(t *testing.T) {
 
 type testBackend struct{}
 
-func (b *testBackend) Config() ml.Config {
+func (b *testBackend) Config() fs.Config {
 	panic("not implemented")
 }
 
diff --git a/ml/backend.go b/ml/backend.go
index cfb18d6a9..b22ba7952 100644
--- a/ml/backend.go
+++ b/ml/backend.go
@@ -9,22 +9,12 @@ import (
 	"slices"
 	"strconv"
 	"strings"
+
+	"github.com/ollama/ollama/fs"
 )
 
-type Config interface {
-	Architecture() string
-	String(string, ...string) string
-	Uint(string, ...uint32) uint32
-	Float(string, ...float32) float32
-	Bool(string, ...bool) bool
-
-	Strings(string, ...[]string) []string
-	Uints(string, ...[]uint32) []uint32
-	Floats(string, ...[]float32) []float32
-}
-
 type Backend interface {
-	Config() Config
+	Config() fs.Config
 	Get(name string) Tensor
 	NewContext() Context
 	NewContextSize(size int) Context
diff --git a/ml/backend/ggml/ggml.go b/ml/backend/ggml/ggml.go
index b6f59ae0e..17f063840 100644
--- a/ml/backend/ggml/ggml.go
+++ b/ml/backend/ggml/ggml.go
@@ -24,7 +24,8 @@ import (
 	"unsafe"
 
 	"github.com/ollama/ollama/format"
-	fs "github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/fs"
+	fsggml "github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/ml"
 	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
 	"golang.org/x/sync/errgroup"
@@ -41,7 +42,7 @@ func devices() []*C.struct_ggml_backend_device {
 }
 
 type Backend struct {
-	meta    *fs.GGML
+	meta    *fsggml.GGML
 	sched   *C.struct_ggml_backend_sched
 	tensors map[string]*C.struct_ggml_tensor
 
@@ -58,7 +59,7 @@ type Backend struct {
 }
 
 func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend, error) {
-	meta, n, err := fs.Decode(r, -1)
+	meta, n, err := fsggml.Decode(r, -1)
 	if err != nil {
 		return nil, err
 	}
@@ -182,7 +183,7 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 	maxTensors += blocks * 2
 
 	type tensor struct {
-		source *fs.Tensor
+		source *fsggml.Tensor
 		target string
 	}
 
@@ -413,7 +414,7 @@ func init() {
 	ml.RegisterBackend("ggml", New)
 }
 
-func (b *Backend) Config() ml.Config {
+func (b *Backend) Config() fs.Config {
 	return b.meta.KV()
 }
 
diff --git a/model/model.go b/model/model.go
index 8355a55a8..bc8944d22 100644
--- a/model/model.go
+++ b/model/model.go
@@ -16,7 +16,8 @@ import (
 	_ "golang.org/x/image/tiff"
 	_ "golang.org/x/image/webp"
 
-	fs "github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/fs"
+	fsggml "github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	_ "github.com/ollama/ollama/ml/backend"
@@ -83,10 +84,10 @@ func (m *Base) Config() config {
 	return m.config
 }
 
-var models = make(map[string]func(ml.Config) (Model, error))
+var models = make(map[string]func(fs.Config) (Model, error))
 
 // Register registers a model constructor for the given architecture
-func Register(name string, f func(ml.Config) (Model, error)) {
+func Register(name string, f func(fs.Config) (Model, error)) {
 	if _, ok := models[name]; ok {
 		panic("model: model already registered")
 	}
@@ -131,14 +132,14 @@ func NewTextProcessor(s string) (TextProcessor, error) {
 		return nil, err
 	}
 	defer r.Close()
-	meta, _, err := fs.Decode(r, -1)
+	meta, _, err := fsggml.Decode(r, -1)
 	if err != nil {
 		return nil, err
 	}
 	return getTextProcessor(meta.KV())
 }
 
-func getTextProcessor(kv fs.KV) (TextProcessor, error) {
+func getTextProcessor(kv fsggml.KV) (TextProcessor, error) {
 	arch := kv.Architecture()
 	f, ok := models[arch]
 	if !ok {
diff --git a/model/model_test.go b/model/model_test.go
index 0b1ea08e8..717c425e2 100644
--- a/model/model_test.go
+++ b/model/model_test.go
@@ -7,7 +7,8 @@ import (
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
-	fs "github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/fs"
+	fsggml "github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/backend/ggml"
 	"github.com/ollama/ollama/ml/nn"
@@ -139,7 +140,7 @@ func TestPopulateFieldsAlternateName(t *testing.T) {
 }
 
 func TestGetTextProcessor(t *testing.T) {
-	tp, err := getTextProcessor(fs.KV{})
+	tp, err := getTextProcessor(fsggml.KV{})
 	if err == nil {
 		t.Error("expected error")
 	} else if !strings.Contains(err.Error(), "unsupported model architecture") {
@@ -148,10 +149,10 @@ func TestGetTextProcessor(t *testing.T) {
 		t.Error("expected nil tp")
 	}
 
-	models["dummy"] = func(ml.Config) (Model, error) {
+	models["dummy"] = func(fs.Config) (Model, error) {
 		return notTextProcessorModel{}, nil
 	}
-	tp, err = getTextProcessor(fs.KV{"general.architecture": "dummy"})
+	tp, err = getTextProcessor(fsggml.KV{"general.architecture": "dummy"})
 	if err == nil {
 		t.Error("expected error")
 	} else if !strings.Contains(err.Error(), "not a TextProcessor") {
diff --git a/model/models/gemma2/model.go b/model/models/gemma2/model.go
index b8f5f0666..752cb5cc2 100644
--- a/model/models/gemma2/model.go
+++ b/model/models/gemma2/model.go
@@ -3,6 +3,7 @@ package gemma2
 import (
 	"math"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
@@ -35,7 +36,7 @@ const (
 	gemma27BLayerCount = 46
 )
 
-func New(c ml.Config) (model.Model, error) {
+func New(c fs.Config) (model.Model, error) {
 	m := Model{
 		SentencePieceModel: model.NewSentencePieceModel(
 			&model.Vocabulary{
diff --git a/model/models/gemma3/model.go b/model/models/gemma3/model.go
index f9c53343a..cef058e2a 100644
--- a/model/models/gemma3/model.go
+++ b/model/models/gemma3/model.go
@@ -6,6 +6,7 @@ import (
 	"math"
 	"slices"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
@@ -52,7 +53,7 @@ func (p *MultiModalProjector) Forward(ctx ml.Context, visionOutputs ml.Tensor, i
 	return visionOutputs
 }
 
-func New(c ml.Config) (model.Model, error) {
+func New(c fs.Config) (model.Model, error) {
 	m := Model{
 		SentencePieceModel: model.NewSentencePieceModel(
 			&model.Vocabulary{
diff --git a/model/models/gemma3/model_text.go b/model/models/gemma3/model_text.go
index 7b2b83c02..3b640a968 100644
--- a/model/models/gemma3/model_text.go
+++ b/model/models/gemma3/model_text.go
@@ -3,6 +3,7 @@ package gemma3
 import (
 	"math"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
@@ -40,7 +41,7 @@ const (
 	cacheTypeCausal
 )
 
-func newTextModel(c ml.Config) *TextModel {
+func newTextModel(c fs.Config) *TextModel {
 	numBlocks := int(c.Uint("block_count"))
 
 	m := TextModel{
diff --git a/model/models/gemma3/model_vision.go b/model/models/gemma3/model_vision.go
index 94aa27bd7..636a363df 100644
--- a/model/models/gemma3/model_vision.go
+++ b/model/models/gemma3/model_vision.go
@@ -3,6 +3,7 @@ package gemma3
 import (
 	"math"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
 )
@@ -111,7 +112,7 @@ func (m *VisionModel) Forward(ctx ml.Context, pixelValues ml.Tensor) ml.Tensor {
 	return hiddenState
 }
 
-func newVisionModel(c ml.Config) *VisionModel {
+func newVisionModel(c fs.Config) *VisionModel {
 	return &VisionModel{
 		Layers: make([]VisionEncoderLayer, c.Uint("vision.block_count")),
 		VisionModelOptions: &VisionModelOptions{
diff --git a/model/models/gemma3/process_image.go b/model/models/gemma3/process_image.go
index fe8269a3b..611a17bd7 100644
--- a/model/models/gemma3/process_image.go
+++ b/model/models/gemma3/process_image.go
@@ -3,7 +3,7 @@ package gemma3
 import (
 	"image"
 
-	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/model/imageproc"
 )
 
@@ -11,7 +11,7 @@ type ImageProcessor struct {
 	imageSize, patchSize, numChannels int
 }
 
-func newImageProcessor(c ml.Config) ImageProcessor {
+func newImageProcessor(c fs.Config) ImageProcessor {
 	return ImageProcessor{
 		imageSize:   int(c.Uint("vision.image_size")),
 		patchSize:   int(c.Uint("vision.patch_size")),
diff --git a/model/models/llama/model.go b/model/models/llama/model.go
index 5c173997b..68980dd76 100644
--- a/model/models/llama/model.go
+++ b/model/models/llama/model.go
@@ -5,6 +5,7 @@ import (
 	"math"
 	"strings"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
@@ -30,7 +31,7 @@ type Model struct {
 	*Options
 }
 
-func New(c ml.Config) (model.Model, error) {
+func New(c fs.Config) (model.Model, error) {
 	if !strings.EqualFold(c.String("tokenizer.ggml.model"), "gpt2") {
 		return nil, fmt.Errorf("tokenizer %s not yet supported", c.String("tokenizer.ggml.model"))
 	}
diff --git a/model/models/mllama/model.go b/model/models/mllama/model.go
index 988a189d4..e53eb184c 100644
--- a/model/models/mllama/model.go
+++ b/model/models/mllama/model.go
@@ -8,6 +8,7 @@ import (
 	"image"
 	"slices"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
@@ -32,7 +33,7 @@ const (
 	selfAttentionLayer
 )
 
-func New(c ml.Config) (model.Model, error) {
+func New(c fs.Config) (model.Model, error) {
 	// Verify unified config
 	if c.Uint("vision.block_count") == 0 {
 		return nil, fmt.Errorf("non-unified vision model not supported")
diff --git a/model/models/mllama/model_text.go b/model/models/mllama/model_text.go
index 1cf30d89b..261897c33 100644
--- a/model/models/mllama/model_text.go
+++ b/model/models/mllama/model_text.go
@@ -4,6 +4,7 @@ import (
 	"math"
 	"slices"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
@@ -220,7 +221,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputIDs, positionIDs, outputs, mask
 	return m.Output.Forward(ctx, hiddenState)
 }
 
-func newTextModel(c ml.Config) *TextModel {
+func newTextModel(c fs.Config) *TextModel {
 	var decoderLayers []TextDecoderLayer
 	for i := range c.Uint("block_count") {
 		var textDecoderLayer TextDecoderLayer
diff --git a/model/models/mllama/model_vision.go b/model/models/mllama/model_vision.go
index ac777f051..2f7d26ca2 100644
--- a/model/models/mllama/model_vision.go
+++ b/model/models/mllama/model_vision.go
@@ -4,6 +4,7 @@ import (
 	"math"
 	"slices"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
 )
@@ -213,7 +214,7 @@ func (m *VisionModel) Forward(ctx ml.Context, pixelValues, positionIDs, aspectRa
 	return hiddenState.Concat(ctx, hiddenStates, 0)
 }
 
-func newVisionModel(c ml.Config) *VisionModel {
+func newVisionModel(c fs.Config) *VisionModel {
 	return &VisionModel{
 		Transformer:       &VisionEncoder{Layers: make([]VisionEncoderLayer, c.Uint("vision.block_count"))},
 		GlobalTransformer: &VisionEncoder{Layers: make([]VisionEncoderLayer, c.Uint("vision.global.block_count"))},
diff --git a/model/models/mllama/process_image.go b/model/models/mllama/process_image.go
index c94d14a64..1b0506d32 100644
--- a/model/models/mllama/process_image.go
+++ b/model/models/mllama/process_image.go
@@ -8,14 +8,14 @@ import (
 
 	"golang.org/x/image/draw"
 
-	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/fs"
 )
 
 type ImageProcessor struct {
 	imageSize, numChannels, maxNumTiles int
 }
 
-func newImageProcessor(c ml.Config) ImageProcessor {
+func newImageProcessor(c fs.Config) ImageProcessor {
 	return ImageProcessor{
 		imageSize:   int(c.Uint("vision.image_size")),
 		numChannels: int(c.Uint("vision.num_channels")),