{"id":23092,"date":"2026-08-07T02:14:29","date_gmt":"2026-08-07T02:14:29","guid":{"rendered":"https:\/\/scannn.com\/introducing-muse-code-and-muse-spark-1-2\/"},"modified":"2026-08-07T02:14:29","modified_gmt":"2026-08-07T02:14:29","slug":"introducing-muse-code-and-muse-spark-1-2","status":"publish","type":"post","link":"https:\/\/scannn.com\/lv\/introducing-muse-code-and-muse-spark-1-2\/","title":{"rendered":"Introducing Muse Code and Muse Spark 1.2"},"content":{"rendered":"\n<div>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">We&#8217;re excited to release Muse\u00a0Code (beta), a terminal coding agent powered by Muse\u00a0Spark\u00a01.2, our newest model. This marks our next step toward the frontier, with larger and much more capable models on the way.<\/p>\n<section class=\"installation-block-module__T-T20q__block\" data-installation-block=\"\">\n<h2 class=\"installation-block-module__T-T20q__heading\">Install Muse\u00a0Code on macOS or Linux:<\/h2>\n<p><button aria-labelledby=\"_R_7pbsnpff5tklb_\" class=\"installation-block-module__T-T20q__copyTarget\" data-copy-state=\"idle\" data-overflow-inline-end=\"false\" type=\"button\"><span class=\"installation-block-module__T-T20q__commandViewport\" data-installation-command-scroll=\"\" data-testid=\"installation-command-scroll\" dir=\"ltr\"><code class=\"installation-block-module__T-T20q__command\" data-language=\"bash\" data-syntax-highlighted=\"true\" data-testid=\"installation-command-code\"><span class=\"installation-block-module__T-T20q__syntaxToken\" data-syntax-token=\"\" data-testid=\"installation-syntax-token\" style=\"--installation-token-dark:#B392F0\">curl<\/span><span class=\"installation-block-module__T-T20q__syntaxToken\" data-syntax-token=\"\" data-testid=\"installation-syntax-token\" style=\"--installation-token-dark:#B392F0\"> <\/span><span class=\"installation-block-module__T-T20q__syntaxToken\" data-syntax-token=\"\" data-testid=\"installation-syntax-token\" style=\"--installation-token-dark:#9DB1C5\">-fsSL<\/span><span class=\"installation-block-module__T-T20q__syntaxToken\" data-syntax-token=\"\" data-testid=\"installation-syntax-token\" style=\"--installation-token-dark:#B392F0\"> <\/span><span class=\"installation-block-module__T-T20q__syntaxToken\" data-syntax-token=\"\" data-testid=\"installation-syntax-token\" style=\"--installation-token-dark:#9DB1C5\">https:\/\/dev.meta.ai\/install.sh<\/span><span class=\"installation-block-module__T-T20q__syntaxToken\" data-syntax-token=\"\" data-testid=\"installation-syntax-token\" style=\"--installation-token-dark:#B392F0\"> <\/span><span class=\"installation-block-module__T-T20q__syntaxToken\" data-syntax-token=\"\" data-testid=\"installation-syntax-token\" style=\"--installation-token-dark:#F97583\">|<\/span><span class=\"installation-block-module__T-T20q__syntaxToken\" data-syntax-token=\"\" data-testid=\"installation-syntax-token\" style=\"--installation-token-dark:#B392F0\"> <\/span><span class=\"installation-block-module__T-T20q__syntaxToken\" data-syntax-token=\"\" data-testid=\"installation-syntax-token\" style=\"--installation-token-dark:#B392F0\">bash<\/span><\/code><\/span><span aria-hidden=\"true\" class=\"installation-block-module__T-T20q__iconFrame\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"24\" height=\"24\" viewbox=\"0 0 24 24\" fill=\"none\" stroke=\"currentColor\" stroke-width=\"1.75\" stroke-linecap=\"round\" stroke-linejoin=\"round\" class=\"lucide lucide-copy installation-block-module__T-T20q__copyIcon\" aria-hidden=\"true\"><rect width=\"14\" height=\"14\" x=\"8\" y=\"8\" rx=\"2\" ry=\"2\"\/><path d=\"M4 16c-1.1 0-2-.9-2-2V4c0-1.1.9-2 2-2h10c1.1 0 2 .9 2 2\"\/><\/svg><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"24\" height=\"24\" viewbox=\"0 0 24 24\" fill=\"none\" stroke=\"currentColor\" stroke-width=\"1.75\" stroke-linecap=\"round\" stroke-linejoin=\"round\" class=\"lucide lucide-check installation-block-module__T-T20q__checkIcon\" aria-hidden=\"true\"><path d=\"M20 6 9 17l-5-5\"\/><\/svg><\/span><span class=\"sr-only\" id=\"_R_7pbsnpff5tklb_\">Copy installation command<\/span><\/button><span aria-live=\"polite\" class=\"sr-only\"\/><\/section>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">Muse\u00a0Code takes on complex software engineering tasks across large repositories: planning changes, writing code, and validating the results. It can coordinate multiple persistent subagents for each task, solving difficult problems faster, more accurately, and with less intervention.<\/p>\n<h2 class=\"font-display text-text-primary pt-8 text-3xl leading-tight font-medium\">Muse\u00a0Code<\/h2>\n<h3 class=\"font-display text-text-primary pt-5 text-xl leading-tight font-medium\">Async Background Agents<\/h3>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">Muse\u00a0Code operates with a simple agent loop plus a set of async background agents to enhance the main agent&#8217;s capability. These specialized background agents remain active throughout each session, rather than being spawned for individual tasks, helping avoid redundant information gathering. They carry out next steps and choose when to communicate back to the main agent. Their persistence reduces latency and the need for steering on difficult, multi-step tasks.<\/p>\n<section aria-label=\"Muse\u00a0Code development examples\" class=\"article-html-demo-tabs-module__kx_NjG__root\" id=\"demos\">\n<div>\n<div class=\"content-tabs-module__Xs1nKW__root\" data-content-tabs=\"\" data-layout=\"tabs\" data-testid=\"content-tabs\">\n<p><span aria-hidden=\"true\" class=\"content-tabs-module__Xs1nKW__selectionIndicator\"\/><button aria-controls=\"_R_4rpbsnpff5tklb_\" aria-selected=\"true\" class=\"content-tabs-module__Xs1nKW__tab\" data-divider=\"hidden\" role=\"tab\" tabindex=\"0\" type=\"button\">Photon Sphere<\/button><button aria-controls=\"_R_4rpbsnpff5tklb_\" aria-selected=\"false\" class=\"content-tabs-module__Xs1nKW__tab\" data-divider=\"hidden\" role=\"tab\" tabindex=\"-1\" type=\"button\">Embervault<\/button><button aria-controls=\"_R_4rpbsnpff5tklb_\" aria-selected=\"false\" class=\"content-tabs-module__Xs1nKW__tab\" data-divider=\"visible\" role=\"tab\" tabindex=\"-1\" type=\"button\">Avo Lawn<\/button><\/p>\n<p><button type=\"button\" role=\"combobox\" aria-expanded=\"false\" aria-autocomplete=\"none\" dir=\"ltr\" data-state=\"closed\" aria-label=\"Muse\u00a0Code development examples\" class=\"content-tabs-module__Xs1nKW__selectTrigger\"><span style=\"pointer-events:none\"\/><span aria-hidden=\"true\" class=\"content-tabs-module__Xs1nKW__selectIcon\"><svg aria-hidden=\"true\" fill=\"none\" focusable=\"false\" height=\"16\" viewbox=\"0 0 16 16\" width=\"16\"><path d=\"M12.8622 5.52875C13.1225 5.26841 13.5445 5.26841 13.8049 5.52875C14.0651 5.78911 14.0652 6.21114 13.8049 6.47146L8.47154 11.8048C8.21122 12.0651 7.78918 12.0651 7.52883 11.8048L2.1955 6.47146C1.93515 6.21112 1.93516 5.78911 2.1955 5.52875C2.45585 5.26842 2.87786 5.26841 3.13821 5.52875L8.00019 10.3907L12.8622 5.52875Z\" fill=\"currentColor\"\/><\/svg><\/span><\/button><select aria-hidden=\"true\" tabindex=\"-1\" style=\"position:absolute;border:0;width:1px;height:1px;padding:0;margin:-1px;overflow:hidden;clip:rect(0, 0, 0, 0);white-space:nowrap;word-wrap:normal\"\/><\/div>\n<\/div>\n<\/section>\n<h3 class=\"font-display text-text-primary pt-5 text-xl leading-tight font-medium\">Runtime Design<\/h3>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">Muse\u00a0Code uses a local event log in which every model call, tool run, approval, and edit is appended. This single source of truth makes the runtime replay-exact and restart-safe: after a crash, the agent can resume precisely where it stopped. That ability lets Muse\u00a0Code take on long-running tasks without being derailed by failures.<\/p>\n<h3 class=\"font-display text-text-primary pt-5 text-xl leading-tight font-medium\">Bundled Skills<\/h3>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">Muse\u00a0Code ships with several default skills. <code class=\"bg-article-inline-code-background rounded-sm px-1 font-mono\">\/plan<\/code> turns a task into an approval-gated plan, <code class=\"bg-article-inline-code-background rounded-sm px-1 font-mono\">\/grill<\/code> stress-tests that plan until it holds up, and <code class=\"bg-article-inline-code-background rounded-sm px-1 font-mono\">\/goal<\/code> works toward successful completion of the specified objective.<\/p>\n<figure class=\"article-media-module__3u0F1q__figure article-media-module__3u0F1q__full\" data-media-placement=\"full\"><figcaption class=\"article-media-module__3u0F1q__caption\">\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">The user inputs a fly-through video of a home into the terminal as an mp4 file. Muse\u00a0Code<br \/>\ninterprets the video and produces a visually rich vacation home marketing and booking page.<\/p>\n<\/figcaption><\/figure>\n<h2 class=\"font-display text-text-primary pt-8 text-3xl leading-tight font-medium\">Muse\u00a0Spark\u00a01.2<\/h2>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">Muse\u00a0Spark\u00a01.2 is a coding-focused update to Muse\u00a0Spark\u00a01.1, with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. In Muse\u00a0Spark\u00a01.2, we significantly scaled up training compute on coding tasks while expanding training environment diversity. The model also maintains its strength in other key areas like general agents.<\/p>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">For more details about our evaluations, see <a class=\"text-link underline underline-offset-4\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/research.meta.ai\/static\/muse-spark-1-2-methodology\">our report<\/a>.<\/p>\n<h3 class=\"font-display text-text-primary pt-5 text-xl leading-tight font-medium\">Co-Training With Muse\u00a0Code<\/h3>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">We co-trained Muse\u00a0Spark\u00a01.2 with Muse\u00a0Code to ensure the model exhibits its best performance and coding usability when paired together. The training included rejection sampled harness trajectories and recipe optimizations for goals, compaction, and subagents, alongside the integration of the Muse\u00a0Code toolset to maximize harness compatibility.<\/p>\n<h3 class=\"font-display text-text-primary pt-5 text-xl leading-tight font-medium\">Long-Horizon<\/h3>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">Muse\u00a0Spark\u00a01.2 was extensively trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects, and auto-research. It leverages planning to sequence work, goal conditioning to maintain direction, and context compaction to retain the knowledge needed to sustain progress.<\/p>\n<h3 class=\"font-display text-text-primary pt-5 text-xl leading-tight font-medium\">Self-Improvement<\/h3>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">We also used Muse\u00a0Spark\u00a01.1 to generate challenging coding environments and instruction-following templates. The model then graded candidate solutions on how well they satisfied those requirements, producing a scalable training dataset for Muse\u00a0Spark\u00a01.2. This self-improvement loop helped Muse\u00a0Spark\u00a01.2 follow complex instructions more precisely than its predecessor.<\/p>\n<h2 class=\"font-display text-text-primary pt-8 text-3xl leading-tight font-medium\">Case Study: Kernel Optimization<\/h2>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">We tested the model&#8217;s ability to iteratively optimize GPU kernels over 1,000+ tool calls (up to 24 hours). Leveraging Muse\u00a0Code&#8217;s agentic coding environment, the model writes, compiles, profiles, and progressively improves kernel performance relative to a provided baseline implementation. We benchmarked on KDA and MLA kernels for NVIDIA Hopper GPUs. The agent continues to achieve substantial improvements over the provided baseline implementation.<\/p>\n<div class=\"article-chart-tabs-module__XJinMW__root\" data-article-chart-tabs=\"\" data-chart-layout=\"stacked\" data-visualization-kind=\"chart\" data-visualization-placement=\"wide\">\n<div class=\"article-chart-tabs-module__XJinMW__panels\">\n<section aria-label=\"KDA\" class=\"article-chart-tabs-module__XJinMW__panel\" data-selected=\"true\" id=\"_R_33pbsnpff5tklb_\" role=\"tabpanel\">\n<div aria-label=\"KDA\" class=\"article-chart-tabs-module__XJinMW__charts\" role=\"group\" style=\"--article-chart-tabs-grid-start:1;--article-chart-tabs-grid-track-count:2\">\n<figure class=\"article-chart-tabs-module__XJinMW__figure\">\n<div class=\"article-chart-tabs-module__XJinMW__mediaShell\" style=\"aspect-ratio:1916 \/ 1080\">\n<div class=\"article-dataset-media-module__b_2lNW__root\" data-article-dataset-media=\"image\"><picture class=\"site-image-module__lM2DSa__picture site-image-module__lM2DSa__fillPicture\" data-site-image-fill=\"true\"><source media=\"(min-width: 64rem)\" srcset=\"https:\/\/research.meta.ai\/articles\/introducing-muse-code-and-muse-spark-1-2\/kernel-optimization\/kda-speedup-v1.png\"\/><\/picture><\/div>\n<\/div>\n<\/figure>\n<\/div>\n<p class=\"article-chart-tabs-module__XJinMW__description\">The baseline is the FLA Triton implementation of KDA. Models were prohibited from importing third-party kernel libraries such as FLA directly; instead, they had to apply specialized kernel-optimization knowledge to implement the algorithm in Triton, rather than wrap existing implementations. Muse\u00a0Spark\u00a01.2 paired a chunk-parallel preparation kernel with a sequential inter-chunk scan, combining standard fusion and tiling with KDA-specific optimizations such as re-centering the gated cumulative decay at the chunk midpoint.<\/p>\n<\/section>\n<section aria-hidden=\"true\" class=\"article-chart-tabs-module__XJinMW__panel\" data-selected=\"false\" role=\"tabpanel\">\n<p class=\"article-chart-tabs-module__XJinMW__description\">We benchmark against a PyTorch reference implementation at batch size 1, number of heads 64, sequence length 8192, and latent dimension 512. Muse\u00a0Spark\u00a01.2 designed a two-kernel Triton pipeline for this workload, combining kernel fusion and tiling with MLA-specific optimizations such as reusing the shared KV latent as both K and V.<\/p>\n<\/section>\n<\/div>\n<\/div>\n<h2 class=\"font-display text-text-primary pt-8 text-3xl leading-tight font-medium\">Availability<\/h2>\n<p class=\"text-text-paragraph text-base leading-7 sm:text-lg sm:leading-8\">Muse\u00a0Spark\u00a01.2 is available today in Muse\u00a0Code and in Meta\u00a0Model\u00a0API with expanded global access. We have a lot on the horizon, including new harness features and more powerful models. We can\u2019t wait to see what you build!<\/p>\n<p><a aria-label=\"Get started with Muse\u00a0Code on the Meta developer site\" class=\"article-promo-module__J04j2W__button\" href=\"https:\/\/dev.meta.ai\" rel=\"noopener noreferrer\" target=\"_blank\">Get started with Muse\u00a0Code<\/a><\/div>\n<p><a href=\"https:\/\/research.meta.ai\/blog\/introducing-muse-code-and-muse-spark-1-2?utm_source=tldrai\">Source link <\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>We&#8217;re excited to release Muse\u00a0Code (beta), a terminal coding agent powered by Muse\u00a0Spark\u00a01.2, our newest model. This marks our next step toward the frontier, with larger and much more capable models on the way. Install Muse\u00a0Code on macOS or Linux: curl -fsSL https:\/\/dev.meta.ai\/install.sh | bashCopy installation command Muse\u00a0Code takes on complex software engineering tasks across [&hellip;]<\/p>\n","protected":false},"author":16,"featured_media":23093,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[143],"tags":[],"class_list":["post-23092","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts\/23092","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/users\/16"}],"replies":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/comments?post=23092"}],"version-history":[{"count":0,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts\/23092\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/media\/23093"}],"wp:attachment":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/media?parent=23092"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/categories?post=23092"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/tags?post=23092"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}