{"id":436,"date":"2026-09-10T22:16:22","date_gmt":"2026-09-10T13:16:22","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/09\/10\/together-ai-open-source-ai-stack-and-might-stack\/"},"modified":"2026-09-20T17:37:19","modified_gmt":"2026-09-20T08:37:19","slug":"together-ai-open-source-ai-stack-and-might-stack","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/10\/together-ai-open-source-ai-stack-and-might-stack\/","title":{"rendered":"The Open-Source AI Stack Explained by Together AI"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Publisher<\/td>\n<td>Together AI<\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code at collection time. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>Together AI has defined the structure of the open-source AI stack as the &#8220;MIGHT stack&#8221; 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.<\/p>\n<h2>Claims and Rationale<\/h2>\n<p>Together AI argues that the AI stack can be broken down into the following five independent layers (the MIGHT stack):<\/p>\n<ul>\n<li><strong>Model<\/strong>: The intelligence layer that interprets requests, performs reasoning, and makes decisions.<\/li>\n<li><strong>Inference<\/strong>: The infrastructure and providers where models actually run.<\/li>\n<li><strong>Gateways and routers<\/strong>: The layer that determines which model or provider handles a request based on cost, speed, and capabilities.<\/li>\n<li><strong>Harness<\/strong>: The application that manages conversations, grants models access to tools, and connects to the codebase.<\/li>\n<li><strong>Tools (Skills and MCP)<\/strong>: Knowledge to execute specific tasks and means to access context.<\/li>\n<\/ul>\n<h3>Model Selection: Large Models vs. Small Models<\/h3>\n<p>The documentation recommends &#8220;model selection as a tool,&#8221; using different model sizes depending on the nature of the task.<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th style=\"text-align: left;\">Feature<\/th>\n<th style=\"text-align: left;\">Large models<\/th>\n<th style=\"text-align: left;\">Small models<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"text-align: left;\"><strong>Main Use Cases<\/strong><\/td>\n<td style=\"text-align: left;\">Complex reasoning, ambiguous tasks, refactoring, understanding the entire codebase<\/td>\n<td style=\"text-align: left;\">Well-defined tasks, test creation, specific error explanation, simple fixes<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><strong>Examples (Model Names)<\/strong><\/td>\n<td style=\"text-align: left;\">Kimi K3 (Total: 1.8T, Active: 104B)<\/td>\n<td style=\"text-align: left;\">GLM 5.3 Flash (Total: 320B, Active: 18B)<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><strong>Advantages<\/strong><\/td>\n<td style=\"text-align: left;\">High robustness, multi-step reasoning, maintaining long conversations<\/td>\n<td style=\"text-align: left;\">Fast response, low cost, high execution speed<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>According to the presenters&#8217; 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.<\/p>\n<h3>Recommended Workflow<\/h3>\n<p>For efficient development, the following workflow called &#8220;Plan, implement, and review&#8221; is proposed:<br \/>\n1. <strong>Plan<\/strong>: Use a large model to break down open-ended prompts into clear tasks.<br \/>\n2. <strong>Implement<\/strong>: Use a small model to execute each task in separate sessions.<br \/>\n3. <strong>Review<\/strong>: Use a large model to review all completed work.<\/p>\n<h2>Prerequisites<\/h2>\n<ul>\n<li>Targets the agentic software development stack using open-weight generative models.<\/li>\n<li>Model options mentioned include GLM 5.3 Flash, DeepSeek V4 Flash, Kimi K3, and MiniMax M3.<\/li>\n<li>Harness examples include PI, OpenCode, and Amp.<\/li>\n<\/ul>\n<h2>What Can Be Replicated Locally<\/h2>\n<p>Readers can build and try their own stack by combining the following components:<\/p>\n<ul>\n<li><strong>Using Inference Providers<\/strong>: Utilize models via APIs through cloud providers such as Together AI.<\/li>\n<li><strong>Introducing Gateways\/Routers<\/strong>:\n<ul>\n<li>Use <a href=\"https:\/\/openrouter.ai\/\">https:\/\/openrouter.ai\/<\/a> or <a href=\"https:\/\/vercel.com\/ai-gateway\">https:\/\/vercel.com\/ai-gateway<\/a> to integrate multiple providers. &#8211; Use <a href=\"https:\/\/www.litellm.ai\/\">https:\/\/www.litellm.ai\/<\/a> to run a router locally or on your own server.<\/li>\n<\/ul>\n<\/li>\n<li><strong>Using Harnesses<\/strong>:\n<ul>\n<li>Utilize existing harnesses such as PI (<a href=\"https:\/\/pi.dev\/\">https:\/\/pi.dev\/<\/a>), OpenCode (<a href=\"https:\/\/opencode.ai\/\">https:\/\/opencode.ai\/<\/a>), and Amp (<a href=\"https:\/\/ampcode.com\/\">https:\/\/ampcode.com\/<\/a>). &#8211; Connect open-weight models to closed harnesses (such as Claude Code) using <code>TogetherLink<\/code>.<\/li>\n<\/ul>\n<\/li>\n<li><strong>Tool Extension<\/strong>:\n<ul>\n<li>Acquire community Skills from <a href=\"https:\/\/www.skills.sh\/\">https:\/\/www.skills.sh\/<\/a>. &#8211; Utilize Model Context Protocol (MCP) servers from <a href=\"https:\/\/mcp.so\/\">https:\/\/mcp.so\/<\/a> to integrate with databases and APIs.<\/li>\n<\/ul>\n<\/li>\n<li><strong>Local Execution<\/strong>: Run models on your own laptop or similar hardware using <code>ollama<\/code>.<\/li>\n<\/ul>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/19\/scaling-coding-agent-traffic-glm-52-together-ai\/\">Scaling Coding Agent Traffic with GLM-5.2 and Dedicated Inference<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/www.together.ai\/blog\/the-open-source-ai-stack\"><a href=\"https:\/\/www.together.ai\/blog\/the-open-source-ai-stack\">https:\/\/www.together.ai\/blog\/the-open-source-ai-stack<\/a><\/a><\/li>\n<\/ul>\n<p><!-- lmw:updates --><\/p>\n<h2>Update History<\/h2>\n<ul>\n<li>2026-09-19: Rewrote the article from re-collected sources and restored it from draft to published.<\/li>\n<\/ul>\n<p><!-- \/lmw:updates --><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Explore Together AI&#8217;s MIGHT stack for open-source AI, detailing layers for models, inference, routing, harnesses, and tools.<\/p>\n","protected":false},"author":1,"featured_media":435,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1528],"tags":[1469,864,1652,1655,866,870,1547],"class_list":["post-436","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-technical-reports","tag-glm-5-3-flash-en","tag-kimi-k3-en","tag-litellm-en","tag-mcp-en","tag-might-stack-en","tag-together-ai-en","tag-verified"],"lang":"en","translations":{"en":436,"ja":434},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/436","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=436"}],"version-history":[{"count":9,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/436\/revisions"}],"predecessor-version":[{"id":2303,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/436\/revisions\/2303"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/435"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=436"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=436"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=436"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}