diff --git a/backend/modules/llm/domain/service/llmimpl/eino/init.go b/backend/modules/llm/domain/service/llmimpl/eino/init.go
index 2688bf13fa..1e50f41fc2 100644
--- a/backend/modules/llm/domain/service/llmimpl/eino/init.go
+++ b/backend/modules/llm/domain/service/llmimpl/eino/init.go
@@ -273,6 +273,10 @@ func ollamaBuilder(ctx context.Context, model *entity.Model, opts ...entity.Opti
 	if err := checkModelBeforeBuild(model); err != nil {
 		return nil, err
 	}
+	return ollama.NewChatModel(ctx, ollamaChatModelConfig(model, opts...))
+}
+
+func ollamaChatModelConfig(model *entity.Model, opts ...entity.Option) *ollama.ChatModelConfig {
 	p := model.ProtocolConfig
 	// 不再从default val里拿，而是从opts里拿这些参数
 	// cp := model.ParamConfig.GetCommonParamDefaultVal()
@@ -285,9 +289,9 @@ func ollamaBuilder(ctx context.Context, model *entity.Model, opts ...entity.Opti
 		Options: &api.Options{
 			// NumKeep:          0,
 			// Seed:             0,
-			// NumPredict:       0,
-			TopK: int(ptr.From(ops.TopK)),
-			TopP: ptr.From(ops.TopP),
+			NumPredict: int(ptr.From(ops.MaxTokens)),
+			TopK:       int(ptr.From(ops.TopK)),
+			TopP:       ptr.From(ops.TopP),
 			// MinP:             0,
 			// TypicalP:         0,
 			// RepeatLastN:      0,
@@ -309,7 +313,7 @@ func ollamaBuilder(ctx context.Context, model *entity.Model, opts ...entity.Opti
 			cfg.KeepAlive = ptr.Of(time.Duration(*pc.KeepAliveMs) * time.Millisecond)
 		}
 	}
-	return ollama.NewChatModel(ctx, cfg)
+	return cfg
 }
 
 func qwenBuilder(ctx context.Context, model *entity.Model, opts ...entity.Option) (einoModel.ToolCallingChatModel, error) {
diff --git a/backend/modules/llm/domain/service/llmimpl/eino/init_ollama_test.go b/backend/modules/llm/domain/service/llmimpl/eino/init_ollama_test.go
new file mode 100644
index 0000000000..4d20a6b610
--- /dev/null
+++ b/backend/modules/llm/domain/service/llmimpl/eino/init_ollama_test.go
@@ -0,0 +1,26 @@
+// Copyright (c) 2025 coze-dev Authors
+// SPDX-License-Identifier: Apache-2.0
+
+package eino
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	"github.com/coze-dev/coze-loop/backend/modules/llm/domain/entity"
+)
+
+func TestOllamaChatModelConfig_MaxTokens(t *testing.T) {
+	model := &entity.Model{ProtocolConfig: &entity.ProtocolConfig{BaseURL: "http://localhost:11434", Model: "llama3"}}
+
+	t.Run("max tokens is passed as num_predict", func(t *testing.T) {
+		cfg := ollamaChatModelConfig(model, entity.WithMaxTokens(256))
+		assert.Equal(t, 256, cfg.Options.NumPredict)
+	})
+
+	t.Run("num_predict stays unset without max tokens", func(t *testing.T) {
+		cfg := ollamaChatModelConfig(model)
+		assert.Equal(t, 0, cfg.Options.NumPredict)
+	})
+}