AI Radar

Muse Spark

Status unknown · Language, Multimodal, Reasoning · Milestone

Muse Spark is the first model in Meta's Muse series, built by Meta Superintelligence Labs. Meta describes it as a natively multimodal reasoning model with tool use; it launched in the Meta AI app and on meta.ai with Instant and Thinking modes, and select partners got a private API preview. [1] [2] Primary source

Timeline of Muse Spark →

Claims and evidence

  • Released Primary source[2] [1]Originally published on April 8, 2026
  • Status Status unknownMeta has no deprecation page. The Model API models page lists only Muse Spark 1.1, 1.2 and 1.3 and calls 1.1 the original version; the Muse Spark 1.1 Evaluation Report says 1.1 succeeds Muse Spark 1.0 as the model powering Meta AI. No Meta statement on the current status of the April 2026 model was found. Primary source[3]Muse Spark; Available Muse Spark models [4]Introduction
  • Replaced by Muse Spark 1.1The report says Muse Spark 1.1 is the updated model powering Meta AI, succeeding Muse Spark 1.0. Primary source[4]Introduction
  • Successor of Llama 4 MaverickThe launch blog calls Llama 4 Maverick our previous model when comparing training compute. The link crosses from the Llama line to the Muse line. Primary source[2]Scaling Axes: Pre-Training
  • Change · Efficiency Meta says a rebuilt pre-training recipe reaches the same capabilities as Llama 4 Maverick with over an order of magnitude less training compute.Compared with Llama 4 Maverick Primary source[2]Scaling Axes: Pre-Training
  • Change · Reasoning Trained with reinforcement learning to reason before answering, with a penalty on thinking time; a Contemplating mode has several agents reason in parallel.Compared with Llama 4 Maverick Primary source[2]Scaling Axes: Test-Time Reasoning
  • Change · Availability Offered in Meta AI and as a private API preview for select partners; Meta said it hoped to open-source future versions.Compared with Llama 4 Maverick Primary source[1]
  • Change · Training data Meta says it worked with more than 1,000 physicians to curate training data for health-related answers.Compared with Llama 4 Maverick Primary source[2]
  • Input text, imageDescribed as natively multimodal; the launch sources describe image understanding (photos, charts) but give no modality list. Primary source[2]first paragraph; Applications: Multimodal [1]Ask Meta AI: It Understands
  • Feature Reasoning modeInstant and Thinking modes in Meta AI; Contemplating mode (several agents reasoning in parallel) rolling out gradually on meta.ai. Primary source[2]first paragraph; Scaling Axes: Test-Time Reasoning; Contemplating mode paragraph [1]Looking Ahead
  • Feature Tool use Primary source[2]first paragraph
  • Parameters Not disclosedNot disclosed by Meta. The newsroom post describes the model as small and fast by design. No claim madeNo source
  • Access consumer appAvailable at launch on meta.ai and in the Meta AI app, rolling out to WhatsApp, Instagram, Facebook, Messenger and AI glasses. Meta also opened a private API preview for select users or partners; api is left out because access was not public. Primary source[2]fourth paragraph [1]Takeaways [5]Abstract

Lineage

Predecessors

Successors

Based on

Not derived from another model.

Variants and derived

Siblings

None recorded.

All ancestors

All descendants

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Meta on AI Radar →