Reapply "don't require pulling stubs for cloud models" again (#14608)

* Revert "Revert "Reapply "don't require pulling stubs for cloud models"" (#14606)"

This reverts commit 39982a954e.

* fix test + do cloud lookup only when seeing cloud models

---------

Co-authored-by: ParthSareen <parth.sareen@ollama.com>

cmd/config/claude.go

@@ -107,15 +107,12 @@ func (c *Claude) ConfigureAliases(ctx context.Context, model string, existingAli
 	}
 
 	if !force && aliases["primary"] != "" {
-		client, _ := api.ClientFromEnvironment()
-		if isCloudModel(ctx, client, aliases["primary"]) {
-			if isCloudModel(ctx, client, aliases["fast"]) {
-				return aliases, false, nil
-			}
-		} else {
-			delete(aliases, "fast")
+		if isCloudModelName(aliases["primary"]) {
+			aliases["fast"] = aliases["primary"]
 			return aliases, false, nil
 		}
+		delete(aliases, "fast")
+		return aliases, false, nil
 	}
 
 	items, existingModels, cloudModels, client, err := listModels(ctx)
@@ -139,10 +136,8 @@ func (c *Claude) ConfigureAliases(ctx context.Context, model string, existingAli
 		aliases["primary"] = primary
 	}
 
-	if isCloudModel(ctx, client, aliases["primary"]) {
-		if aliases["fast"] == "" || !isCloudModel(ctx, client, aliases["fast"]) {
-			aliases["fast"] = aliases["primary"]
-		}
+	if isCloudModelName(aliases["primary"]) {
+		aliases["fast"] = aliases["primary"]
 	} else {
 		delete(aliases, "fast")
 	}

cmd/config/config.go

@@ -233,6 +233,9 @@ func ModelExists(ctx context.Context, name string) bool {
 	if name == "" {
 		return false
 	}
+	if isCloudModelName(name) {
+		return true
+	}
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return false

cmd/config/droid.go

@@ -10,7 +10,6 @@ import (
 	"path/filepath"
 	"slices"
 
-	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 )
 
@@ -125,13 +124,12 @@ func (d *Droid) Edit(models []string) error {
 	}
 
 	// Build new Ollama model entries with sequential indices (0, 1, 2, ...)
-	client, _ := api.ClientFromEnvironment()
 
 	var newModels []any
 	var defaultModelID string
 	for i, model := range models {
 		maxOutput := 64000
-		if isCloudModel(context.Background(), client, model) {
+		if isCloudModelName(model) {
 			if l, ok := lookupCloudModelLimit(model); ok {
 				maxOutput = l.Output
 			}

cmd/config/droid_test.go

@@ -1276,35 +1276,17 @@ func TestDroidEdit_LocalModelDefaultMaxOutput(t *testing.T) {
 
 func TestDroidEdit_CloudModelLimitsUsed(t *testing.T) {
 	// Verify that every cloud model in cloudModelLimits has a valid output
-	// value that would be used for maxOutputTokens when isCloudModel returns true.
-	// Cloud suffix normalization must also work since integrations may see either
-	// :cloud or -cloud model names.
+	// value that would be used for maxOutputTokens when the selected model uses
+	// the explicit :cloud source tag.
 	for name, expected := range cloudModelLimits {
 		t.Run(name, func(t *testing.T) {
-			l, ok := lookupCloudModelLimit(name)
-			if !ok {
-				t.Fatalf("lookupCloudModelLimit(%q) returned false", name)
-			}
-			if l.Output != expected.Output {
-				t.Errorf("output = %d, want %d", l.Output, expected.Output)
-			}
-			// Also verify :cloud suffix lookup
 			cloudName := name + ":cloud"
-			l2, ok := lookupCloudModelLimit(cloudName)
+			l, ok := lookupCloudModelLimit(cloudName)
 			if !ok {
 				t.Fatalf("lookupCloudModelLimit(%q) returned false", cloudName)
 			}
-			if l2.Output != expected.Output {
-				t.Errorf(":cloud output = %d, want %d", l2.Output, expected.Output)
-			}
-			// Also verify -cloud suffix lookup
-			dashCloudName := name + "-cloud"
-			l3, ok := lookupCloudModelLimit(dashCloudName)
-			if !ok {
-				t.Fatalf("lookupCloudModelLimit(%q) returned false", dashCloudName)
-			}
-			if l3.Output != expected.Output {
-				t.Errorf("-cloud output = %d, want %d", l3.Output, expected.Output)
+			if l.Output != expected.Output {
+				t.Errorf("output = %d, want %d", l.Output, expected.Output)
 			}
 		})
 	}

cmd/config/integrations.go

@@ -14,6 +14,7 @@ import (
 
 	"github.com/ollama/ollama/api"
 	internalcloud "github.com/ollama/ollama/internal/cloud"
+	"github.com/ollama/ollama/internal/modelref"
 	"github.com/ollama/ollama/progress"
 	"github.com/spf13/cobra"
 )
@@ -326,12 +327,7 @@ func SelectModelWithSelector(ctx context.Context, selector SingleSelector) (stri
 
 	// If the selected model isn't installed, pull it first
 	if !existingModels[selected] {
-		if cloudModels[selected] {
-			// Cloud models only pull a small manifest; no confirmation needed
-			if err := pullModel(ctx, client, selected); err != nil {
-				return "", fmt.Errorf("failed to pull %s: %w", selected, err)
-			}
-		} else {
+		if !isCloudModelName(selected) {
 			msg := fmt.Sprintf("Download %s?", selected)
 			if ok, err := confirmPrompt(msg); err != nil {
 				return "", err
@@ -526,7 +522,7 @@ func selectModelsWithSelectors(ctx context.Context, name, current string, single
 
 	var toPull []string
 	for _, m := range selected {
-		if !existingModels[m] {
+		if !existingModels[m] && !isCloudModelName(m) {
 			toPull = append(toPull, m)
 		}
 	}
@@ -552,12 +548,28 @@ func selectModelsWithSelectors(ctx context.Context, name, current string, single
 	return selected, nil
 }
 
+// TODO(parthsareen): consolidate pull logic from call sites
 func pullIfNeeded(ctx context.Context, client *api.Client, existingModels map[string]bool, model string) error {
-	if existingModels[model] {
+	if isCloudModelName(model) || existingModels[model] {
 		return nil
 	}
-	msg := fmt.Sprintf("Download %s?", model)
-	if ok, err := confirmPrompt(msg); err != nil {
+	return confirmAndPull(ctx, client, model)
+}
+
+// TODO(parthsareen): pull this out to tui package
+// ShowOrPull checks if a model exists via client.Show and offers to pull it if not found.
+func ShowOrPull(ctx context.Context, client *api.Client, model string) error {
+	if _, err := client.Show(ctx, &api.ShowRequest{Model: model}); err == nil {
+		return nil
+	}
+	if isCloudModelName(model) {
+		return nil
+	}
+	return confirmAndPull(ctx, client, model)
+}
+
+func confirmAndPull(ctx context.Context, client *api.Client, model string) error {
+	if ok, err := confirmPrompt(fmt.Sprintf("Download %s?", model)); err != nil {
 		return err
 	} else if !ok {
 		return errCancelled
@@ -569,26 +581,6 @@ func pullIfNeeded(ctx context.Context, client *api.Client, existingModels map[st
 	return nil
 }
 
-// TODO(parthsareen): pull this out to tui package
-// ShowOrPull checks if a model exists via client.Show and offers to pull it if not found.
-func ShowOrPull(ctx context.Context, client *api.Client, model string) error {
-	if _, err := client.Show(ctx, &api.ShowRequest{Model: model}); err == nil {
-		return nil
-	}
-	// Cloud models only pull a small manifest; skip the download confirmation
-	// TODO(parthsareen): consolidate with cloud config changes
-	if strings.HasSuffix(model, "cloud") {
-		return pullModel(ctx, client, model)
-	}
-	if ok, err := confirmPrompt(fmt.Sprintf("Download %s?", model)); err != nil {
-		return err
-	} else if !ok {
-		return errCancelled
-	}
-	fmt.Fprintf(os.Stderr, "\n")
-	return pullModel(ctx, client, model)
-}
-
 func listModels(ctx context.Context) ([]ModelItem, map[string]bool, map[string]bool, *api.Client, error) {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -733,10 +725,8 @@ func syncAliases(ctx context.Context, client *api.Client, ac AliasConfigurer, na
 	}
 	aliases["primary"] = model
 
-	if isCloudModel(ctx, client, model) {
-		if aliases["fast"] == "" || !isCloudModel(ctx, client, aliases["fast"]) {
-			aliases["fast"] = model
-		}
+	if isCloudModelName(model) {
+		aliases["fast"] = model
 	} else {
 		delete(aliases, "fast")
 	}
@@ -1022,7 +1012,7 @@ Examples:
 				existingAliases = aliases
 
 				// Ensure cloud models are authenticated
-				if isCloudModel(cmd.Context(), client, model) {
+				if isCloudModelName(model) {
 					if err := ensureAuth(cmd.Context(), client, map[string]bool{model: true}, []string{model}); err != nil {
 						return err
 					}
@@ -1211,7 +1201,7 @@ func buildModelList(existing []modelInfo, preChecked []string, current string) (
 	// When user has no models, preserve recommended order.
 	notInstalled := make(map[string]bool)
 	for i := range items {
-		if !existingModels[items[i].Name] {
+		if !existingModels[items[i].Name] && !cloudModels[items[i].Name] {
 			notInstalled[items[i].Name] = true
 			var parts []string
 			if items[i].Description != "" {
@@ -1305,7 +1295,8 @@ func IsCloudModelDisabled(ctx context.Context, name string) bool {
 }
 
 func isCloudModelName(name string) bool {
-	return strings.HasSuffix(name, ":cloud") || strings.HasSuffix(name, "-cloud")
+	// TODO(drifkin): Replace this wrapper with inlining once things stabilize a bit
+	return modelref.HasExplicitCloudSource(name)
 }
 
 func filterCloudModels(existing []modelInfo) []modelInfo {

cmd/config/integrations_test.go

@@ -426,8 +426,14 @@ func TestBuildModelList_NoExistingModels(t *testing.T) {
 	}
 
 	for _, item := range items {
-		if !strings.HasSuffix(item.Description, "(not downloaded)") {
-			t.Errorf("item %q should have description ending with '(not downloaded)', got %q", item.Name, item.Description)
+		if strings.HasSuffix(item.Name, ":cloud") {
+			if strings.HasSuffix(item.Description, "(not downloaded)") {
+				t.Errorf("cloud model %q should not have '(not downloaded)' suffix, got %q", item.Name, item.Description)
+			}
+		} else {
+			if !strings.HasSuffix(item.Description, "(not downloaded)") {
+				t.Errorf("item %q should have description ending with '(not downloaded)', got %q", item.Name, item.Description)
+			}
 		}
 	}
 }
@@ -492,10 +498,14 @@ func TestBuildModelList_ExistingRecommendedMarked(t *testing.T) {
 			if strings.HasSuffix(item.Description, "(not downloaded)") {
 				t.Errorf("installed recommended %q should not have '(not downloaded)' suffix, got %q", item.Name, item.Description)
 			}
-		case "minimax-m2.5:cloud", "kimi-k2.5:cloud", "qwen3:8b":
+		case "qwen3:8b":
 			if !strings.HasSuffix(item.Description, "(not downloaded)") {
 				t.Errorf("non-installed recommended %q should have '(not downloaded)' suffix, got %q", item.Name, item.Description)
 			}
+		case "minimax-m2.5:cloud", "kimi-k2.5:cloud":
+			if strings.HasSuffix(item.Description, "(not downloaded)") {
+				t.Errorf("cloud model %q should not have '(not downloaded)' suffix, got %q", item.Name, item.Description)
+			}
 		}
 	}
 }
@@ -536,7 +546,13 @@ func TestBuildModelList_HasRecommendedCloudModel_OnlyNonInstalledAtBottom(t *tes
 	}
 
 	for _, item := range items {
-		if !slices.Contains([]string{"kimi-k2.5:cloud", "llama3.2"}, item.Name) {
+		isCloud := strings.HasSuffix(item.Name, ":cloud")
+		isInstalled := slices.Contains([]string{"kimi-k2.5:cloud", "llama3.2"}, item.Name)
+		if isInstalled || isCloud {
+			if strings.HasSuffix(item.Description, "(not downloaded)") {
+				t.Errorf("installed or cloud model %q should not have '(not downloaded)' suffix, got %q", item.Name, item.Description)
+			}
+		} else {
 			if !strings.HasSuffix(item.Description, "(not downloaded)") {
 				t.Errorf("non-installed %q should have '(not downloaded)' suffix, got %q", item.Name, item.Description)
 			}
@@ -1000,8 +1016,8 @@ func TestShowOrPull_ModelNotFound_ConfirmNo_Cancelled(t *testing.T) {
 	}
 }
 
-func TestShowOrPull_CloudModel_SkipsConfirmation(t *testing.T) {
-	// Confirm prompt should NOT be called for cloud models
+func TestShowOrPull_CloudModel_DoesNotPull(t *testing.T) {
+	// Confirm prompt should NOT be called for explicit cloud models
 	oldHook := DefaultConfirmPrompt
 	DefaultConfirmPrompt = func(prompt string) (bool, error) {
 		t.Error("confirm prompt should not be called for cloud models")
@@ -1032,8 +1048,115 @@ func TestShowOrPull_CloudModel_SkipsConfirmation(t *testing.T) {
 	if err != nil {
 		t.Errorf("ShowOrPull should succeed for cloud model, got: %v", err)
 	}
-	if !pullCalled {
-		t.Error("expected pull to be called for cloud model without confirmation")
+	if pullCalled {
+		t.Error("expected pull not to be called for cloud model")
+	}
+}
+
+func TestShowOrPull_CloudLegacySuffix_DoesNotPull(t *testing.T) {
+	// Confirm prompt should NOT be called for explicit cloud models
+	oldHook := DefaultConfirmPrompt
+	DefaultConfirmPrompt = func(prompt string) (bool, error) {
+		t.Error("confirm prompt should not be called for cloud models")
+		return false, nil
+	}
+	defer func() { DefaultConfirmPrompt = oldHook }()
+
+	var pullCalled bool
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/api/show":
+			w.WriteHeader(http.StatusNotFound)
+			fmt.Fprintf(w, `{"error":"model not found"}`)
+		case "/api/pull":
+			pullCalled = true
+			w.WriteHeader(http.StatusOK)
+			fmt.Fprintf(w, `{"status":"success"}`)
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	defer srv.Close()
+
+	u, _ := url.Parse(srv.URL)
+	client := api.NewClient(u, srv.Client())
+
+	err := ShowOrPull(context.Background(), client, "gpt-oss:20b-cloud")
+	if err != nil {
+		t.Errorf("ShowOrPull should succeed for cloud model, got: %v", err)
+	}
+	if pullCalled {
+		t.Error("expected pull not to be called for cloud model")
+	}
+}
+
+func TestPullIfNeeded_CloudModel_DoesNotPull(t *testing.T) {
+	oldHook := DefaultConfirmPrompt
+	DefaultConfirmPrompt = func(prompt string) (bool, error) {
+		t.Error("confirm prompt should not be called for cloud models")
+		return false, nil
+	}
+	defer func() { DefaultConfirmPrompt = oldHook }()
+
+	err := pullIfNeeded(context.Background(), nil, map[string]bool{}, "glm-5:cloud")
+	if err != nil {
+		t.Fatalf("expected no error for cloud model, got %v", err)
+	}
+
+	err = pullIfNeeded(context.Background(), nil, map[string]bool{}, "gpt-oss:20b-cloud")
+	if err != nil {
+		t.Fatalf("expected no error for cloud model with legacy suffix, got %v", err)
+	}
+}
+
+func TestSelectModelsWithSelectors_CloudSelection_DoesNotPull(t *testing.T) {
+	var pullCalled bool
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/api/status":
+			w.WriteHeader(http.StatusNotFound)
+			fmt.Fprintf(w, `{"error":"not found"}`)
+		case "/api/tags":
+			w.WriteHeader(http.StatusOK)
+			fmt.Fprintf(w, `{"models":[]}`)
+		case "/api/me":
+			w.WriteHeader(http.StatusOK)
+			fmt.Fprintf(w, `{"name":"test-user"}`)
+		case "/api/pull":
+			pullCalled = true
+			w.WriteHeader(http.StatusOK)
+			fmt.Fprintf(w, `{"status":"success"}`)
+		default:
+			w.WriteHeader(http.StatusNotFound)
+			fmt.Fprintf(w, `{"error":"not found"}`)
+		}
+	}))
+	defer srv.Close()
+	t.Setenv("OLLAMA_HOST", srv.URL)
+
+	single := func(title string, items []ModelItem, current string) (string, error) {
+		for _, item := range items {
+			if item.Name == "glm-5:cloud" {
+				return item.Name, nil
+			}
+		}
+		t.Fatalf("expected glm-5:cloud in selector items, got %v", items)
+		return "", nil
+	}
+
+	multi := func(title string, items []ModelItem, preChecked []string) ([]string, error) {
+		return nil, fmt.Errorf("multi selector should not be called")
+	}
+
+	selected, err := selectModelsWithSelectors(context.Background(), "codex", "", single, multi)
+	if err != nil {
+		t.Fatalf("selectModelsWithSelectors returned error: %v", err)
+	}
+	if !slices.Equal(selected, []string{"glm-5:cloud"}) {
+		t.Fatalf("unexpected selected models: %v", selected)
+	}
+	if pullCalled {
+		t.Fatal("expected cloud selection to skip pull")
 	}
 }
 

cmd/config/opencode.go

@@ -12,8 +12,8 @@ import (
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/internal/modelref"
 )
 
 // OpenCode implements Runner and Editor for OpenCode integration
@@ -26,14 +26,13 @@ type cloudModelLimit struct {
 }
 
 // lookupCloudModelLimit returns the token limits for a cloud model.
-// It normalizes common cloud suffixes before checking the shared limit map.
+// It normalizes explicit cloud source suffixes before checking the shared limit map.
 func lookupCloudModelLimit(name string) (cloudModelLimit, bool) {
-	// TODO(parthsareen): migrate to using cloud check instead.
-	for _, suffix := range []string{"-cloud", ":cloud"} {
-		name = strings.TrimSuffix(name, suffix)
-	}
-	if l, ok := cloudModelLimits[name]; ok {
-		return l, true
+	base, stripped := modelref.StripCloudSourceTag(name)
+	if stripped {
+		if l, ok := cloudModelLimits[base]; ok {
+			return l, true
+		}
 	}
 	return cloudModelLimit{}, false
 }
@@ -150,8 +149,6 @@ func (o *OpenCode) Edit(modelList []string) error {
 		}
 	}
 
-	client, _ := api.ClientFromEnvironment()
-
 	for _, model := range modelList {
 		if existing, ok := models[model].(map[string]any); ok {
 			// migrate existing models without _launch marker
@@ -161,7 +158,7 @@ func (o *OpenCode) Edit(modelList []string) error {
 					existing["name"] = strings.TrimSuffix(name, " [Ollama]")
 				}
 			}
-			if isCloudModel(context.Background(), client, model) {
+			if isCloudModelName(model) {
 				if l, ok := lookupCloudModelLimit(model); ok {
 					existing["limit"] = map[string]any{
 						"context": l.Context,
@@ -175,7 +172,7 @@ func (o *OpenCode) Edit(modelList []string) error {
 			"name":    model,
 			"_launch": true,
 		}
-		if isCloudModel(context.Background(), client, model) {
+		if isCloudModelName(model) {
 			if l, ok := lookupCloudModelLimit(model); ok {
 				entry["limit"] = map[string]any{
 					"context": l.Context,

cmd/config/opencode_test.go

@@ -714,16 +714,17 @@ func TestLookupCloudModelLimit(t *testing.T) {
 		wantContext int
 		wantOutput  int
 	}{
-		{"glm-4.7", true, 202_752, 131_072},
+		{"glm-4.7", false, 0, 0},
 		{"glm-4.7:cloud", true, 202_752, 131_072},
 		{"glm-5:cloud", true, 202_752, 131_072},
 		{"gpt-oss:120b-cloud", true, 131_072, 131_072},
 		{"gpt-oss:20b-cloud", true, 131_072, 131_072},
-		{"kimi-k2.5", true, 262_144, 262_144},
+		{"kimi-k2.5", false, 0, 0},
 		{"kimi-k2.5:cloud", true, 262_144, 262_144},
-		{"deepseek-v3.2", true, 163_840, 65_536},
+		{"deepseek-v3.2", false, 0, 0},
 		{"deepseek-v3.2:cloud", true, 163_840, 65_536},
-		{"qwen3-coder:480b", true, 262_144, 65_536},
+		{"qwen3-coder:480b", false, 0, 0},
+		{"qwen3-coder:480b:cloud", true, 262_144, 65_536},
 		{"qwen3-coder-next:cloud", true, 262_144, 32_768},
 		{"llama3.2", false, 0, 0},
 		{"unknown-model:cloud", false, 0, 0},