Skip to content
FrontierDot

Meta Muse Model Timeline: Muse Spark to Muse Wonder and Wonder Full, with Prices (2026)

Updated

Live timeline · last updated · 2 entries in the update log

Meta Superintelligence Labs shipped seven Muse releases in about six months. Here is each one with its date and Meta’s own source, plus what Muse Wonder costs, how Muse Wonder Full differs, and what Meta has not shown.

As of Oct 24, 2026 Meta’s newest Muse models are Muse Wonder and the subscription-only Muse Wonder Full, launched Oct 12, 2026, and the open-weight Muse Spark Open from Oct 15; on Oct 22 Paradigm, Meta and Andromeda opened a $100,000 H200 kernel challenge for that open build.

Read on 2026-10-21 for the model line, and again on 2026-10-24 (UTC+8) for the kernel challenge: Meta’s Muse Spark, Muse Spark 1.1 and Muse Spark 1.3 posts, the Muse Code and Muse Wonder developer posts, the Muse Spark Open announcement, Hugging Face’s launch post, and Paradigm’s challenge post. Benchmark tables on Meta’s pages are Meta’s own runs; we did not reproduce them.

Timeline

  1. Meta introduces Muse Spark, the first model in the Muse family from Meta Superintelligence Labs: a natively multimodal reasoning model with tool use, visual chain of thought and multi-agent orchestration. Meta calls it the first step on its scaling ladder.

    Meta AI: Introducing Muse Spark

  2. Muse Spark 1.1 is released as a significant upgrade, with gains in tool and computer use, coding and multimodal understanding. Developers can use it through the Meta Model API, now in public preview.

    Meta AI: Introducing Muse Spark 1.1

  3. Muse Glimmer, a small dense Muse model meant for running locally, is released this month according to Hugging Face. We found no official page with a day, so this node carries only the month.

    Hugging Face: Welcome Muse Spark Open

  4. Meta releases Muse Spark 1.2 and an early beta of Muse Code, a coding agent for long-horizon, multi-agent work whose subagents, tool calls, steers and cancels are recorded in a replayable event log.

    Meta for Developers: Build with Muse Code

  5. Muse Spark 1.3 becomes available in Muse Code and the Meta Model API. Meta says it uses about 20% fewer tool calls and about 25% fewer tokens.

    Meta Research: Introducing Muse Spark 1.3

  6. Meta launches Muse Wonder on Muse Code and the Meta Model API at $4 input, $0.20 cached input and $20 output per million tokens, with a 1M-token context. Muse Wonder Full, Meta’s most advanced model, is available only by subscription in Muse Code on the Power Usage plan and a new Maximum Usage plan with 80x the Everyday Usage limit.

    Meta for Developers: Put your ideas to work with Muse Wonder

  7. Meta releases the weights of Muse Spark Open, an 816B-total, 42B-active model with a 1M-token context under the Muse Model License.

    Meta Research: Introducing Muse Spark Open

  8. Paradigm, with Meta and Andromeda, launches the Muse Spark Open Kernel Challenge: a $100,000 prize pool to speed Muse Spark Open on H200 GPUs, with kernel and serving tracks ending November 19.

    Paradigm: Muse Spark Open Kernel Challenge

Muse Wonder vs Muse Wonder Full

The two share a name and a 1M-token context window but are sold differently. Muse Wonder is a metered model: $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, available in Muse Code and through the Meta Model API. Meta says it brings performance comparable to Muse Wonder Full.

Muse Wonder Full has no per-token price on the launch page. It is available only through a Muse Code subscription, on the Power Usage plan or the new Maximum Usage plan, which Meta says gives 80x the Everyday Usage limit with the same five-hour and weekly resets. If you want to call Wonder Full from your own code through the API, the launch page does not offer that.

What Meta says Wonder is for

Meta positions the Muse Wonder series for coding agents, long-running agents and computer use. The launch post lists multi-step reasoning, end-to-end agentic task completion, multi-file refactors across large repos, recovery from tool-call failures, and multimodal reasoning over screenshots, charts, documents, images and video. It adds two features it calls active memory, which carries context across conversations, and active focus, which trains the model to try alternative approaches and to ask for clarification before important actions.

The post also says Muse Wonder Full tops the DeepSWE and SWE-Marathon coding leaderboards in independent third-party runs, and it publishes a table against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5. We report that this claim exists; we have not checked the third-party runs ourselves.

One detail of that table is worth knowing before you quote it. Every column header is labeled “(max)”, meaning maximum effort, for Muse Wonder Full, Muse Wonder and the three competing models alike. Results at the default or lower effort settings most applications use can differ, and the table does not show cost per task, so it says little about which model is cheaper for a given job.

The Spark releases, one by one

Muse Spark on April 8 was Meta’s reset. The announcement calls it the first product of a ground-up overhaul of Meta’s AI efforts and the first step on a scaling ladder, and describes it as natively multimodal with tool use, visual chain of thought and multi-agent orchestration. It did not come with a developer API price.

Muse Spark 1.1 on July 9 is where developers came in. Meta called it a significant upgrade, with major gains in tool and computer use, coding and multimodal understanding, and made it available through the Meta Model API in public preview. The same post ties it to the launch of Muse Image that week.

Muse Spark 1.2 on August 5 shipped together with the first beta of Muse Code, Meta’s coding agent. The developer post stresses auditability: every subagent, tool call, steer and cancel is observable and replayable through an event log. Muse Spark 1.3 on September 2 followed in Muse Code and the Meta Model API. Meta says it uses about 20% fewer tool calls and about 25% fewer tokens, and that it improved adversarial robustness and resistance to prompt injections on agentic tasks.

Muse Spark Open on October 15 made weights public for the first time. Our Muse Spark Open post covers it in detail; the PyTorch blog says the open version shares the same base model as Muse Spark 1.3.

The current Muse lineup

Meta’s developer site lists these models in its Muse menu as of October 21, 2026: Muse Wonder, Muse Spark 1.3, Muse Spark Open, Muse Glimmer, Muse Image and Muse Voice, alongside Transcribe and Segment Anything (SAM 3.1). This timeline covers the language models; image, voice and segmentation models are left out.

For developers choosing between them, the practical split is simple. Muse Wonder is the metered frontier option through the API, Muse Wonder Full is for Muse Code subscribers, Muse Spark 1.3 is the earlier model still offered in Muse Code and the API, and Muse Spark Open and Muse Glimmer are the options you can download and run yourself. Check Meta’s pricing pages for any model other than Muse Wonder, since we only verified Wonder’s price.

How fast the Muse line moved

  • Muse Spark to Muse Spark 1.1: about three months (April 8 to July 9).
  • Spark 1.1 to Spark 1.2: under four weeks (July 9 to August 5).
  • Spark 1.2 to Spark 1.3: four weeks (August 5 to September 2).
  • Spark 1.3 to Muse Wonder: under six weeks (September 2 to October 12).
  • Muse Wonder to Muse Spark Open weights: three days (October 12 to October 15).

What is not verified

  • Parameter counts for Muse Spark, Muse Wonder or Muse Wonder Full. Meta has published them only for Muse Spark Open (816B total, 42B active).
  • An exact day for Muse Glimmer. Our only source gives the month.
  • Whether Muse Wonder Full will reach the Meta Model API. The launch page does not say.
  • Benchmark results beyond Meta’s own table.

Related reading

For the open-weight release, read Muse Spark Open: 816B weights. To compare with other vendors, see the Claude model timeline and the OpenAI GPT model timeline. The Front blog index lists every post.

The same format also covers Moonshot Kimi, Mistral, DeepSeek and Alibaba Qwen.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

When was Muse Wonder released?

October 12, 2026, on Muse Code and the Meta Model API.

How much does Muse Wonder cost?

$4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, according to Meta’s launch page.

How do I get Muse Wonder Full?

Through a Muse Code subscription on the Power Usage or Maximum Usage plan. Meta’s launch page lists no API price for it.

What was the first Muse model?

Muse Spark, introduced by Meta Superintelligence Labs on April 8, 2026.

Update log

  • — Added the Oct 22, 2026 Muse Spark Open Kernel Challenge node.
  • — First published. Checked Meta’s Muse announcements through Oct 21, 2026 (UTC+8).

Sources