Muse Spark 1.3: Cheap Contributor Tokens Are Not a Private-Data Tier

What Meta launched and who it fits

On September 2, 2026, Meta released Muse Spark 1.3 as an agentic and coding update, rolling it out in Muse Code and the Meta Model API. If you run long tool-calling sessions or use a hosted model with mixed-media context, it is worth your attention.

If you were hoping to download it, stop here. Meta’s launch material describes open weights as future work, and a public Hugging Face search for the exact model name returned nothing on 2026-09-02. There is no local path for Apple silicon, RTX, or DGX Spark. This is a cloud model, so every decision below assumes your prompts leave your machine. The question is a boundary, not hardware: Meta has shipped two tiers with sharply different terms.

The real decision: Standard versus Contributor

Muse Spark 1.3 is listed on OpenRouter under two model IDs: meta/muse-spark-1.3 and meta/muse-spark-1.3-contributor. OpenRouter calls the second tier Contributor; Meta’s terms call it Discounted Services. They look like a pricing choice. They are a data-use choice with a price attached.

For Standard service, Meta says content is not used to train Meta Models. Read that narrowly. It does not say zero retention, and it does not say no human review. Meta states it may still process content for abuse and terms enforcement and for safety and security systems, and it describes retention that may apply in some circumstances. Standard means “not fed to model training,” not “nobody will ever see this.”

For the discounted Contributor service, Meta says inputs and outputs may be used to train, develop, evaluate, and improve Meta AI models. Meta also indicates that some evaluation, safety, abuse, quality, and policy review can occur before its stated disassociation steps.

The operational consequence is blunt. On Contributor, do not send:

• private or proprietary repositories
• credentials, tokens, keys, or connection strings
• customer data, user records, or anything covered by a data processing agreement
• client work you were paid to keep confidential
• persistent agent memory files, which quietly accumulate all of the above

That last one catches people. Point a long-lived agent memory or scratchpad at the cheap tier and you are not sharing one prompt — you are sharing the compressed history of everything that agent has ever handled. Contributor is not a discount on Standard. It is a different deal in which your work is part of the payment.

What you actually get (context, modalities, tools, provider paths)

The capability surface is identical across both tiers; only the terms differ. As listed on OpenRouter on 2026-09-02, both model IDs carry a 1M-token context window. Inputs include text, images, video, files and PDFs, and audio. Output is text. Tool calling and structured outputs are supported, which is what makes it usable as an agent backend rather than a chat toy.

One caveat from the provider itself: OpenRouter warns that audio understanding is not fully supported and audio quality may degrade. Treat audio as experimental and verify it by hand before building a workflow on it.

On access paths, Meta lists Muse Code and the Meta Model API. OpenRouter offers both tiers and lists Meta as the single upstream, forwarding requests directly. Routing through a gateway does not create an independent data-processing boundary. Send Contributor traffic through a third party and the Contributor terms still apply at the far end. Meta also publishes a Muse Code install command for macOS and Linux in the launch post; use the vendor’s instructions and confirm the current model identifier before wiring anything up.

Price, with a dated and volatile warning

Rates listed on OpenRouter on 2026-09-02:

Tier	Input / M	Cache read / M	Output / M
Standard	$1.25	$0.15	$4.25
Contributor	$0.10	$0.002	$0.20

These are volatile. A launch-week table is a snapshot, not a contract — re-check the live provider page before you build a cost model or commit a budget.

Read the gap honestly. Contributor input is roughly an order of magnitude cheaper and output roughly twenty times cheaper. That deserves a policy decision rather than a default. The price gap reflects a material terms difference.

What Meta’s performance claims do and do not prove

Meta reports that its engineers measured approximately 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 in its comparisons.

That is vendor-reported evidence from vendor-designed comparisons. It is a reasonable prompt to run your own test; it is not a guaranteed saving on your workload, and it has not been independently verified. Meta’s evaluation methodology and launch scorecard are its own materials, and the reasoning modes available to you may not match the conditions behind the published numbers. ABS has not tested this model. If efficiency is your reason for switching, count tool calls and tokens on a task you already run, before and after, and let your numbers decide.

A safe first evaluation

The goal of a first run is to learn something without spending anything you can’t get back.

1. Pick a disposable task — synthetic, public, or already published. Nothing you would mind seeing in a training set.
2. Start on Standard if the task touches anything real. Use Contributor only after deliberately classifying the work as shareable or disposable.
3. Confirm the exact model identifier on the provider page you are actually calling.
4. Run an agent task you already use, unchanged, and log tool calls, total tokens, and wall-clock time as a baseline.
5. Test tool calling and structured outputs explicitly, since those break first in real harnesses.
6. Only then decide which tier each workflow belongs on, and write that decision down where your team can see it.

That last step is the one people skip. The failure mode is not choosing wrong once — it is having no rule, so the cheap tier becomes the default for whatever someone runs late on a Friday.

Known boundaries and when not to use it

No local or offline option. Open weights are future work per Meta’s launch material, and no download path was found in public sources on 2026-09-02. If you need on-device inference or air-gapped operation, this release is not a candidate — look at open-weight models instead.

No published quotas here. No source-backed current rate limits for Muse Spark 1.3 were retrieved. Check your live account tier and the provider’s own documentation instead of trusting a number from an article — including this one.

Standard is not an enterprise agreement. Meta describes TLS 1.2 and above in transit, access controls, and account content deletion controls, while noting some retention may apply. Standard is the minimum sensible tier for private context, but where you are contractually required to hold a written retention or processing agreement, a public commitments page is not a substitute.

Preview terms are moving. The Meta Model API terms were last updated August 28, 2026 and describe a limited preview period. Preview means the surface can change under you.

No peer comparison here. There is no fair apples-to-apples benchmark evidence against other frontier models, so this guide makes no ranking claim.

Done means

You are finished when all of these are true:

• You know which tier each workflow uses, and why.
• No credential, customer record, private repository, client deliverable, or persistent agent memory can reach the Contributor tier — checked, not assumed.
• You have run one disposable evaluation task and logged tool calls and tokens against your current model.
• You have re-checked live pricing on the provider page rather than relying on the table above.
• Your tier rule is written down where a teammate would find it before making the same choice.

What this article does NOT cover

• Local installation, quantization, or hardware sizing — there is no released local checkpoint.
• Direct Meta API endpoint tutorials, authentication flows, or account setup steps.
• Numeric rate limits, RPM or TPM quotas, or throughput guarantees.
• Available reasoning modes or maximum-reasoning configuration.
• Head-to-head benchmarks against other vendors’ models, or any ABS-run testing of this one.

Related guides

• Muse Spark 1.2 — the earlier release in this line.
• Muse Glimmer-30B — Meta’s open-weight sibling model.
• How to find, download, and evaluate local models on Hermes or OpenClaw — start here if what you need is a model you can run yourself.

Sources

• Meta AI Research — Introducing Muse Spark 1.3
• Meta — AI developer data commitments
• Meta — Model API terms of service
• OpenRouter — meta/muse-spark-1.3 (Standard)
• OpenRouter — meta/muse-spark-1.3-contributor
• Meta — Muse Spark 1.3 multimodal evaluation methodology

Research dossier: Johnny5/Websites/Agentic Botsitting/research/muse-spark-1-3/muse-spark-1-3-research-2026-09-02.md