Changelog

Shipping in the open.

Every release of the Muse Code CLI and Meta’s Muse models, as it lands.

Read Meta’s Muse Wonder post →
muse-spark-openOct 15, 2026OPEN

Muse Spark Open: open weights for the Spark line

NEWOpen-weight MoE model: 816B total parameters, 42B active, 1M-token context; text, image, video, and audio in, text out
NEWMuse Model License (custom): free to run, fine-tune, and redistribute, with extra terms for model-serving businesses over $20M in revenue
NEWWeights on Hugging Face, SageMaker, Google Cloud, and Microsoft Foundry; hosted by Together AI, Fireworks, Groq, OpenRouter, and others
NEWNeeds about 700 GB of accelerator memory (NVFP4); Meta demoed Muse Code on it across two M5 Ultra Mac Studios
muse-wonderOct 12, 2026MODEL

Muse Wonder and Wonder Full arrive in Muse Code

NEWMuse Wonder, Meta’s first frontier model, in Muse Code and on the Model API as muse-wonder-1.0 at $4 / $20 per 1M tokens (no Contributor variant)
NEWMuse Wonder Full for Muse Code subscribers on Power and Maximum plans; not available on the API
NEWMaximum Usage plan: $200 a month, 80× the Everyday allowance
FIXEDLaunch-day bug that counted Wonder usage against plan limits too fast; Meta said it was fixed and usage reset
muse code 1.4.xSep–Oct 2026CLI

CLI 1.4 series: Windows, plugins, images, and muse serve

NEWNative Windows install from PowerShell: irm https://dev.meta.ai/install.ps1 | iex
NEW/model <model-id>, automatic session naming, /resume <id-or-name>, and import of Claude or Codex skills
NEWPlugins and marketplaces on by default; PostToolBatch and StopFailure hooks
NEWVideo attachments, image generation and editing, and /blender for rebuilding reference images as Blender scenes
IMPROVEDContributor mode on by default for subscribers; fixed background requests that still used it after it was turned off
IMPROVEDEnterprise settings: ca_bundle for corporate certificates, session_retention_days, and -c key=value overrides
muse code plansSep 2026PLANS

Bigger allowances on High and Power

IMPROVEDHigh Usage raised from 3× to 5× the Everyday allowance, and Power Usage from 10× to 20×, at the same prices (no announcement; visible on Meta’s plan page by September 21)
muse-spark-1.3Sep 2, 2026MODEL

Muse Spark 1.3 becomes Muse Code’s model

NEWReplaces Muse Spark 1.2 in Muse Code; same-day release on the Meta Model API as muse-spark-1.3 and muse-spark-1.3-contributor
IMPROVEDAbout 20% fewer tool calls and 25% fewer tokens than 1.2 on the same coding work
IMPROVEDAsks clarifying questions and flags when it is stuck instead of guessing
NEWPrices unchanged: $1.25 / $4.25 per 1M tokens on Standard, $0.10 / $0.20 on Contributor
NEWMax reasoning configuration in limited partner preview while safety testing finishes
muse codeAug 31, 2026GA

Muse Code is out of beta

NEWMonthly plans: Everyday Usage ($5), High Usage ($15, 3× usage), Power Usage ($50, 10× usage); pay-as-you-go stays
NEWInter-session messaging: sessions on the same machine warn and answer each other over a local socket
NEWWorkflows: orchestrate teams of subagents across stages, monitored live in /workflows
NEWRewind: press Esc twice to roll the conversation back to a safe point from the event log
NEWMuse Code SDK in developer preview: TypeScript, built on the local Muse Session Protocol
muse-glimmerAug 10, 2026OPEN

Muse Glimmer: an open-weight sibling

NEWOpen-weight 30B agentic model distilled from Muse Spark, released under Apache 2.0
NEWRuns locally on a single consumer GPU (about 20 GB quantized); weights on Hugging Face, Ollama, and LM Studio
NEWA separate model: Meta has not documented it as a Muse Code backend
muse code betaAug 5, 2026BETA

Muse Code beta × Muse Spark 1.2

NEWMuse Code beta for macOS and Linux — one-command install, browser auth at dev.meta.ai
NEWPersistent async background agents — active all session, not spawned per task
NEWAgent fan-out: sub-agents work in parallel in isolated git worktrees
NEWAppend-only local event log — replay-exact, restart-safe after crashes
NEWBuilt-in /plan, /grill, and /goal commands
NEWZero-data-retention requests now accepted via Meta sales
muse-spark-1.2Aug 5, 2026MODEL

Muse Spark 1.2 on Meta Model API

NEWCo-trained with the Muse Code harness — rejection-sampled agent trajectories, goal conditioning, context compaction
IMPROVEDOptimized for multi-file refactors, long debugging sessions, and tasks that run past a single prompt
IMPROVEDKernel case study: 1,000+ tool calls sustained over 24 hours
NEWContributor and standard variants; expanded global access to the public preview
muse-spark-1.1Earlier in 2026

Muse Spark 1.1 public preview

NEWMeta Model API opens: OpenAI-SDK-compatible endpoint at api.meta.ai/v1
NEW1,048,576-token context window
NEWModel API cookbook: API fundamentals, agent patterns, and end-to-end use cases