Meta

Muse Spark

Provider

Meta

Status

available

Context

1,000,000 tok

SWE-bench

55%

Price

$1.25 / $4.25 /MTok

Muse Spark is the first model from Meta’s Meta Superintelligence Labs, launched on 8 April 2026 and spearheaded by chief AI officer Alexandr Wang. It is the clearest signal of Meta’s strategy reset: a proprietary, natively multimodal reasoning model that replaced the open Llama models as the engine behind Meta AI — a deliberate break from the open-weights approach Meta built its AI reputation on (CNBC).

Meta describes it as “small and fast by design,” and the efficiency claim is the headline: Meta says Muse Spark matches Llama 4 Maverick with over 10x less compute (Artificial Analysis). It is a capable first-generation frontier-adjacent model — independent evaluator Artificial Analysis scores it 52 on its Intelligence Index v4.0, 4th overall at release — that leads on health questions and vision, but sits below the very top of the field.

Verdict — updated 4 September 2026. Muse Spark is very much a live, evolving line, now on its fourth release in five months. On 9 July Meta shipped Muse Spark 1.1 — an agentic upgrade with a 1M-token context — and opened the Meta Model API, the first time it has charged for its own model. On 5 August it followed with Muse Spark 1.2 and Muse Code, a terminal coding agent, plus a $0.10/$0.20 data-contributor pricing tier that is roughly a twelfth of the standard rate; on 10 August it released Muse Glimmer — a 30B Apache 2.0 open-weights model distilled from 1.2. On 3 September came Muse Spark 1.3, an agentic-coding point release whose strongest numbers — including the 62 Artificial Analysis measured, above GPT-6 Astra’s 61 — come from a max configuration still in limited preview. So: Muse Spark is Meta’s current assistant and coding engine, aggressively priced, still closed-weights — Meta’s promised open weights remain unversioned and undated — and its shipping configuration is still unscored by any independent index.

Quick specs

TRY MUSE SPARK →

Muse Spark 1.3 (3 September 2026)

Meta shipped Muse Spark 1.3 on 3 September 2026 through Muse Code and the Meta Model API — the fourth Muse Spark release in five months. It is an agentic-coding point release: Meta claims 75.4% on DeepSWE v1.1 against 1.2’s 59.3%, completing the same tasks with roughly 20% fewer tool calls and 25% fewer tokens. Pricing is unchanged: $1.25/$4.25 standard, $0.10/$0.20 on the data-contributor tier.

Read the launch table carefully. Meta’s strongest 1.3 numbers come from its max reasoning configuration, which is still completing safety testing and available only in a limited partner preview — Meta says it will arrive “shortly”. That preview config is also what Artificial Analysis scored: 62 on the Intelligence Index, which would place it above GPT-6 Astra (61) and level with Claude Fable 5 — but you cannot buy that configuration today, and no independent score exists for the standard model that ships. Our board credits the shipping configuration, which is why Muse Spark 1.3 holds at #22 rather than jumping into the flagship pack; if max ships generally and the score holds, that placement gets revisited upward.

Zuckerberg posted the same day that Muse Spark open-weights releases are coming “soon” — still with no version, date, model size or licence attached.

Muse Spark 1.2 and Muse Code (5 August 2026)

Meta shipped Muse Spark 1.2 on 5 August 2026 alongside Muse Code, a terminal-based coding agent in beta — its answer to Claude Code and OpenAI’s Codex. Read the release correctly: 1.2 is a coding-focused point release of 1.1, not a new base model, and it remains closed-weights with the same 1M-token context.

The pricing is the real news. The standard rate is unchanged at $1.25 in / $4.25 out per million tokens, but Meta added a data-contributor tier at $0.10 in / $0.20 out — roughly a twelfth of the standard price — where your prompts and outputs may be used to improve Meta products. The standard tier carries no such usage. That is an unusually explicit “pay with data or pay with money” split, and at $0.10/$0.20 it undercuts nearly everything at comparable capability.

Meta’s coding claims are not yet verifiable. Meta reports 82.9% on Terminal-Bench 2.1 for Muse Spark 1.2 driving Muse Code, and 59.3% on DeepSWE — but the public Terminal-Bench and DeepSWE boards did not list Muse Spark 1.2 when checked on 7 August 2026, so there is no like-for-like public entry to compare against GPT-5.6 Sol or Claude Opus 5. Independent testing from Vals is more sober: 14th on common-harness Terminal-Bench, but 5th overall on the Vals Index at about $0.69 per test. The honest read is the one that follows from the price: a capable, very cheap coding model — not a benchmark leader.

Five days later Meta distilled 1.2 into the open-weights Muse Glimmer (30B, Apache 2.0) and said open weights for Muse Spark 1.2 itself are coming — which, if it happens, would matter considerably more than either release so far.

What’s new in Muse Spark 1.1

Meta upgraded the line to Muse Spark 1.1 on 9 July 2026, and the release did two notable things.

It went agentic. Muse Spark 1.1 is built for agentic work — multi-step tool use, computer use and coding — with reported gains over the original across tool use, computer use, coding and multimodal understanding, a 1M-token context (up from ~262K) and multi-agent orchestration. In Meta AI that shows up as an assistant that “doesn’t just think, it acts”: it can make plans, connect to email and calendar, build slides and carry out tasks on your behalf (Meta). Meta framed it explicitly as an entry into the AI coding market, chasing Anthropic and OpenAI (CNBC).

It became Meta’s first paid model. Alongside 1.1, Meta opened the Meta Model API in public preview — the first time outside developers can build on a Muse model, and the first time Meta has charged for its own model, at roughly $1.25 in / $4.25 out per million tokens (AI Weekly). Muse Spark 1.1 is closed, hosted and metered per token — a structural reversal of the open-weights Llama strategy Meta built its AI reputation on.

The wider Muse family. Two days earlier, on 7 July 2026, Meta Superintelligence Labs shipped its first media-generation models: Muse Image (now in Meta AI, generating and blending images with reasoning) and a Muse Video preview (Meta). Together with 1.1, this burst of releases is what drove the model’s mid-2026 coverage.

What Muse Spark is

Muse Spark is the first in a planned “Muse” family from Meta Superintelligence Labs, the organisation Mark Zuckerberg built around Alexandr Wang after a $14 billion-plus deal for a stake in Scale AI (CNBC). It is natively multimodal — accepting text, image and speech input — with tool use, visual chain-of-thought and an Instant / Thinking mode split for quick answers versus deeper reasoning (Meta).

The strategically important fact is that it is proprietary. Where Llama defined Meta’s open-weights era, Muse Spark keeps its weights closed (Meta says it “hopes to open-source future versions”), prioritising integration into Meta’s own products and ecosystem (Artificial Analysis). See the Meta provider page for the wider pivot.

Benchmark performance

Vendor- and Artificial-Analysis-reported figures for the original April model; treat as a ceiling pending broader independent testing. Muse Spark 1.1 (9 July 2026) targets agentic, coding and multimodal gains over these numbers, but independent 1.1 benchmarks had not landed at the time of writing.

Muse Spark’s profile is “strong generalist, standout on health and vision, efficient” rather than frontier-leading (Artificial Analysis). Its Contemplating mode orchestrates multiple agents reasoning in parallel, letting it compete with the extreme-reasoning modes of frontier models (≈58% on Humanity’s Last Exam). On our best AI models ranking it sits in the Strong tier.

Capabilities

Muse Spark handles text, image and speech input with text output, a 1M-token context window (up from ~262K in the original April model), tool use, visual chain-of-thought and multi-agent orchestration (the Contemplating mode, and 1.1’s agentic tool/computer use). The efficiency story is central: Meta claims it reaches Llama 4 Maverick-level capability with more than ten times less compute (Artificial Analysis), consistent with the “small and fast by design” framing.

Access and pricing

Muse Spark is free for general users on meta.ai and the Meta AI app, where it powers the assistant across Meta’s products. For developers, the picture changed with 1.1: the Meta Model API entered public preview on 9 July 2026 at roughly $1.25 in / $4.25 out per million tokens — Meta’s first metered, closed, hosted model access (AI Weekly).

Muse Spark 1.2 (5 August) kept that rate and added a second option:

Muse Spark itself remains proprietary with no open-weights release — a contrast with Llama 4 — though the distilled Muse Glimmer is now open under Apache 2.0, and Meta says open weights for Muse Spark 1.2 are coming.

How Muse Spark compares

• vs the frontier — Muse Spark trails Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro on the aggregate Intelligence Index, but leads them on HealthBench and is near the top on vision.
• vs Llama 4 — Muse Spark replaced Llama as the Meta AI engine and matches Llama 4 Maverick’s capability at a fraction of the compute, but it is closed where Llama is open.
• vs Muse Glimmer — Glimmer is the distilled, open, local sibling: 30B parameters under Apache 2.0, a 131K context, and free to self-host on a 24GB consumer GPU. Muse Spark 1.2 is far more capable with a 1M-token context, but it is closed, hosted and metered. Choose Spark for ceiling, Glimmer for control.

Known limitations

Below the frontier leaders on overall capability, and independent benchmarks for 1.1 and 1.2 remain thin — Meta’s 1.2 coding claims (82.9% Terminal-Bench 2.1, 59.3% DeepSWE) had no matching public-board entry when checked. Still proprietary and metered — no open weights for Muse Spark itself, though the distilled Muse Glimmer is now Apache 2.0 and Meta says 1.2 weights are coming. The data-contributor tier has a real cost: the $0.10/$0.20 rate is cheap because your prompts and outputs may be used to improve Meta products, which rules it out for confidential work. And the wider Meta AI privacy posture (ad personalisation from chats) applies — see Meta AI.

Frequently asked questions

What is Muse Spark 1.3?

Muse Spark 1.3 is the current version, released 3 September 2026 through Muse Code and the Meta Model API at unchanged pricing. Meta claims 75.4% on DeepSWE v1.1 (up from 1.2’s 59.3%) with roughly 20% fewer tool calls and 25% fewer tokens — but its strongest numbers, including the 62 Artificial Analysis measured on its Intelligence Index, come from a max configuration still in limited preview and safety testing. No independent score yet exists for the configuration you can actually use.

What is Muse Spark?

Muse Spark is the first model from Meta Superintelligence Labs, launched 8 April 2026. It is a proprietary, natively multimodal reasoning model that replaced Llama as the engine behind Meta AI, and Meta’s first major step away from open weights.

How good is Muse Spark?

Capable but not frontier-leading. Artificial Analysis scores it 52 on its Intelligence Index (4th overall at release), behind Gemini 3.1 Pro, GPT-5.4 and Claude Opus 4.6 — but it leads on HealthBench Hard and is among the best vision models, and is notably compute-efficient.

Is Muse Spark open source?

No. Unlike Llama, Muse Spark is proprietary — Meta has not released its weights, though it says open weights for Muse Spark 1.2 are coming. What Meta has released is Muse Glimmer, a 30B model distilled from 1.2 and published under Apache 2.0 on 10 August 2026. Muse Spark itself remains free in Meta’s apps and metered through the public-preview Meta Model API.

How much does Muse Spark cost?

It is free for consumers in the Meta AI app and at meta.ai. For developers, Muse Spark 1.2 is available through the Meta Model API at $1.25 in / $4.25 out per million tokens on the standard tier, or $0.10 in / $0.20 out on the data-contributor tier, where your prompts and outputs may be used to improve Meta products.

What is Muse Spark 1.2?

Muse Spark 1.2, released 5 August 2026, is the current version — a coding-focused point release of 1.1 rather than a new base model, shipped alongside Muse Code, a terminal-based coding agent in beta. It introduced the $0.10/$0.20 data-contributor pricing tier. Meta reports 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE, but those figures were not listed on the public boards, and independent Vals testing places it 14th on common-harness Terminal-Bench while ranking it 5th overall on the Vals Index for value at roughly $0.69 per test.

What is Muse Code?

Muse Code is Meta’s terminal-based coding agent, released in beta on 5 August 2026 alongside Muse Spark 1.2 — Meta’s entry into the same category as Claude Code and OpenAI’s Codex CLI. Meta’s headline Terminal-Bench figure (82.9%) is for Muse Spark 1.2 driving Muse Code, not the model alone.

What is Muse Spark 1.1?

Muse Spark 1.1, released 9 July 2026, is the current version of the model. It’s an agentic upgrade — stronger tool use, computer use, coding and multimodal understanding, with a 1M-token context and multi-agent orchestration — and it launched alongside the Meta Model API, the first time Meta has charged developers for its own model. In Meta AI it powers an assistant that can plan, connect to email and calendar, build slides and complete multi-step tasks.

Last verified 12 August 2026. Headline benchmark figures are Meta- and Artificial-Analysis-reported from the April 2026 launch; Muse Spark 1.1 (9 July) and 1.2 (5 August) improve on them, but independent numbers for both remain thin and Meta’s 1.2 coding claims were not listed on the public Terminal-Bench or DeepSWE boards when checked. Confirm against Meta’s official pages before relying on specific numbers.