SearchBook a call

BULLETIN / Meta

Muse Spark 1.3: review

Meta's closed Muse Spark 1.3 is fast and cheap for its quality, but it is new to developers and its weights are not yet released.

A low-priced closed model with solid independent scores, a young developer platform and no open weights yet.

Maker
Meta
Released
2 Sept 2026
Licence
Proprietary
Access
Meta Model API, Muse Code, OpenRouter
Context
1M tokens
Price
$1.25 input / $4.25 output per million tokens (USD); contributor tier $0.10 / $0.20

Good at

  • Quality for the price
  • Output speed
  • Image and video understanding
  • Long context of 1M tokens

Bad at

  • Young API and tooling
  • No published parameter count or detailed benchmark table
  • Behind the leaders on maths
  • No open weights despite Llama history

Use it for

  • Cost-sensitive agent and coding work
  • Multimodal apps with image and video input
  • Teams wanting a third vendor for redundancy

Avoid it for

  • Anyone who needs to self-host (see Muse Glimmer)
  • Sensitive data on the cheap contributor tier
  • Work that needs the very top scores

Yes, if you want good results per dollar and do not need to run the model yourself. Muse Spark 1.3 scores below the Anthropic and OpenAI leaders on independent tests, 48 against 58 on the Artificial Analysis index, at a much lower price. It is not an open model: Meta’s current flagship is closed.

First principles

  • A closed model is rented. You buy access, not the model. Your prompts go to the vendor’s servers, and the vendor decides when the model changes or is retired. Open weights reverse this: you carry the hardware cost and keep control. Muse Spark 1.3 is the first kind. Meta’s open option today is the much smaller Muse Glimmer.
  • A cheaper tier can be paid for in data. The contributor tier costs a fraction of the standard price, and Meta marks the tier as used to improve its products. On our reading, part of the price is your data, so the tier suits public or synthetic data only.
  • With little disclosure, independent tests are the evidence. Meta gives no parameter count, no architecture and no benchmark numbers we could tie to named tests. This review therefore rests on Artificial Analysis, the Vals Index, Epoch AI and Arena, and reports no Meta benchmark scores.
  • A second vendor is insurance. Any single vendor can raise prices, change a model’s behaviour or go down. A model that is good enough and priced low is worth wiring in as a fallback, even where it is not your first choice.

What it is

Muse Spark 1.3 is the current flagship from Meta Superintelligence Labs, released on 2 September 2026. Meta says it improves agentic and coding tasks over version 1.2, using about 20% fewer tool calls and about 25% fewer tokens on coding. It takes text, images, video and documents as input and returns text. The context window is 1M tokens. Meta has not published the parameter count.

It is available through the Meta Model application programming interface (API) at dev.meta.ai, in Meta’s Muse Code terminal tool, and on OpenRouter.

The Muse Spark line began on 8 April 2026. Meta then said the first version had gaps in long-horizon agent work and coding.

Meta’s open model today is a separate family, Muse Glimmer. Muse-Glimmer-30B, released in August 2026 under Apache 2.0, has about 29.6B parameters and takes text and images. Meta says it needs 64GB of video memory at full precision, or 24 to 32GB in quantised form. Before Muse, Meta’s open models were the Llama family; Llama 4 Maverick was released in April 2025.

Where it is good

Value. On the Vals Index (updated 29 September 2026), Muse Spark 1.3 at max effort scored 58.16% at $3.78 per test. That puts it seventh, ahead of GPT-6 Sol (57.54% at $7.58) and Grok 4.7 (54.95% at $12.12). Only Anthropic’s models and GPT-6 Astra scored higher.

Speed and index score. Artificial Analysis gave it 48 at max effort, with about 174 output tokens per second and a cost of $1.60 per index task. That is faster than any Claude or GPT-6 entry near it.

Vision and web. On the Arena vision board, Muse Spark models hold three places in the top ten, with version 1.3 at max effort tenth on 1290. On the Arena web development board (updated 29 September 2026) it ranks eleventh at 1655.

Price. The standard tier is $1.25 input and $4.25 output per million tokens, with cached input at $0.15.

Where it is bad

It trails the leaders on hard maths and science. Epoch AI recorded 74.4% at xhigh effort on FrontierMath Tiers 1 to 3, against 93.7% for GPT-6 Astra and 91.2% for Claude Opus 5.5. Its Epoch Capabilities Index of 156.9 sits below Gemini 3.8 Flash and Kimi K3.

Transparency is limited. Meta’s announcement shows a benchmark chart but the page text gives no numbers we could tie to named tests, so we report none from Meta. Parameter count and architecture are undisclosed.

The cheap tier has a cost of its own. The contributor tier, at $0.10 input and $0.20 output, is marked as used to improve Meta’s products. On our reading of the page, that means your requests may feed Meta’s development. It is unsuitable for confidential data.

There are no open weights. Meta’s 1.3 announcement mentions an open weights release on its roadmap, with no date. We could not confirm from a Meta page whether version 1.3 powers the consumer Meta AI app.

The older open line has restrictive terms. The Llama 4 licence requires a separate licence for companies with more than 700 million monthly active users, and its use policy withholds some rights to the multimodal models from users based in the European Union (EU).

Price and access

Tier Input per million Cached input Output per million
Standard $1.25 $0.15 $4.25
Contributor $0.10 $0.002 $0.20

Who should use it

On this evidence it is a strong pick for cost-sensitive agent, coding and multimodal work, and a good second vendor to reduce dependence on one lab. Use the standard tier for anything private. Teams that need to host a Meta model themselves should look at Muse Glimmer instead, and accept a much smaller model.

Sources