Skip to main content
Some Tensormesh serverless reasoning models expose reasoning controls through chat template settings. The provider forwards model-specific options unchanged through providerOptions.tensormesh.chat_template_kwargs.

​
DeepSeek-V4-Flash

For the default non-thinking mode, omit reasoning controls: 
import { generateText } from "ai";
import { tensormesh } from "@tensormesh/ai-sdk-provider";

const { text } = await generateText({
  model: tensormesh("deepseek-ai/DeepSeek-V4-Flash"),
  prompt: "What is 17*19? Return only the final integer.",
  temperature: 1,
  topP: 1,
});
For thinking modes, pass thinking: true and the model’s reasoning_effort inside chat_template_kwargs: 
const { text } = await generateText({
  model: tensormesh("deepseek-ai/DeepSeek-V4-Flash"),
  prompt:
    "Find the three-digit lock code. The digits are all different. The digits sum to 14. The first digit is twice the third digit. The second digit is one less than the first digit. Return only the code.",
  temperature: 1,
  topP: 1,
  providerOptions: {
    tensormesh: {
      chat_template_kwargs: {
        thinking: true,
        reasoning_effort: "high",
      },
    },
  },
});
Use reasoning_effort: "max" when you specifically want DeepSeek-V4-Flash Think Max.

​
AI SDK Reasoning Fields

Use providerOptions.tensormesh.chat_template_kwargs for Tensormesh chat-template reasoning controls. The AI SDK/OpenAI-compatible reasoning and reasoningEffort settings are standard request fields and are sent as top-level reasoning_effort, not inside chat_template_kwargs.