The Open-Source AI Stack Explained by Together AI

September 10, 2026September 20, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

Together AI has defined the structure of the open-source AI stack as the “MIGHT stack" and explains the role of each layer and the flexibility of their combinations. It demonstrates how developers transitioning from closed models to open-weight models should select and customize the model, inference, gateway, harness, and tool layers.

Claims and Rationale

Together AI argues that the AI stack can be broken down into the following five independent layers (the MIGHT stack):

• Model: The intelligence layer that interprets requests, performs reasoning, and makes decisions.
• Inference: The infrastructure and providers where models actually run.
• Gateways and routers: The layer that determines which model or provider handles a request based on cost, speed, and capabilities.
• Harness: The application that manages conversations, grants models access to tools, and connects to the codebase.
• Tools (Skills and MCP): Knowledge to execute specific tasks and means to access context.

Model Selection: Large Models vs. Small Models

The documentation recommends “model selection as a tool," using different model sizes depending on the nature of the task.

According to the presenters’ comparison, GLM 5.3 Flash is reported to be about 6 times smaller in parameter count and about 20 times cheaper in cost compared to Kimi K3.

Recommended Workflow

For efficient development, the following workflow called “Plan, implement, and review" is proposed:
1. Plan: Use a large model to break down open-ended prompts into clear tasks.
2. Implement: Use a small model to execute each task in separate sessions.
3. Review: Use a large model to review all completed work.

Prerequisites

• Targets the agentic software development stack using open-weight generative models.
• Model options mentioned include GLM 5.3 Flash, DeepSeek V4 Flash, Kimi K3, and MiniMax M3.
• Harness examples include PI, OpenCode, and Amp.

What Can Be Replicated Locally

Readers can build and try their own stack by combining the following components:

• Using Inference Providers: Utilize models via APIs through cloud providers such as Together AI.
• Introducing Gateways/Routers:
  • Use https://openrouter.ai/ or https://vercel.com/ai-gateway to integrate multiple providers. – Use https://www.litellm.ai/ to run a router locally or on your own server.
• Using Harnesses:
  • Utilize existing harnesses such as PI (https://pi.dev/), OpenCode (https://opencode.ai/), and Amp (https://ampcode.com/). – Connect open-weight models to closed harnesses (such as Claude Code) using TogetherLink.
• Tool Extension:
  • Acquire community Skills from https://www.skills.sh/. – Utilize Model Context Protocol (MCP) servers from https://mcp.so/ to integrate with databases and APIs.
• Local Execution: Run models on your own laptop or similar hardware using ollama.

Related Articles

• Scaling Coding Agent Traffic with GLM-5.2 and Dedicated Inference

Sources

• https://www.together.ai/blog/the-open-source-ai-stack

Update History

• 2026-09-19: Rewrote the article from re-collected sources and restored it from draft to published.