Meta AI Mac
Current engine behind Meta AI

Not an upgrade. A trade.

Muse Spark 1.3 became the model behind Meta AI on 2 September 2026, with a one-million-token context window and real gains on tool use. It also got measurably worse at one thing: reading long documents accurately. That makes it the rare release where the honest summary is not “better” but “better at some of your work and worse at the rest”, and which half you live in decides how you feel about it.

Which work it helps The published figures
1M token context 4 published figures 4 kinds of work 1 measured regression
Tool-use benchmarks rising from 1.2 to 1.3 while long-document accuracy falls 1.2 → 1.3, ON THE THREE FIGURES WITH A BEFORE AND AFTER 35% 47% Tau3-Bench Banking · Meta 80% 85% Terminal-Bench 2.1 Meta 83% 79% Long documents independent testing Faded bars are 1.2. Two arrows point up, one points down, and the one pointing down was not measured by Meta. Whether that is an upgrade depends on what you do with it
Bars are drawn to a common baseline so the moves are comparable. The gains are real; so is the loss.
Features, as published

What Meta shipped, and who measured what

Four figures describe this release. Three came from Meta and one did not, and that distinction is worth carrying with you rather than flattening into a single impression.

Meta’s figure

Tau3-Bench Banking

35% 47% multi-step tool use

The largest single move in the release. In plain terms the model completes about a third more of these tasks than 1.2 did, on a benchmark built around chaining tool calls to finish a job.

Meta’s figure

Terminal-Bench 2.1

80% 85% command-line work

A smaller headline move that matters more than it looks, because it is measured against a high base. Read as failures rather than scores, roughly a quarter of the remaining errors went away.

Meta’s figure

Tool calls per task

~20% fewer efficiency

Meta reports the model reaching answers in fewer steps. Users experience this as briskness and, as the section below argues, possibly as something less welcome too.

Independent testing

Long-document accuracy

83% 79% the regression

Not a Meta figure, and the only one that moved the wrong way. In failure terms it is about a quarter more errors, which is close to the mirror image of the Terminal-Bench gain.

The million-token window and the awkward pairing

1.3 also carries a one-million-token context window that accepts text, images and video enough to hold a very large document, or a stack of them, in one session. Set beside the regression, that produces an odd result: the window grew large enough to hold the documents the model became slightly worse at reading carefully.

There is no contradiction in it. Capacity and attention are different properties, and being able to fit something in context has never been the same as reliably noticing everything inside it. It is a useful corrective to the idea that a bigger window is automatically a better reader.

Our reading of it

Three descriptions that may be one change

Fewer tool calls, feeling more literal, and slipping on long documents are usually reported as three separate facts about 1.3. They are also exactly what you would expect from a single underlying shift.

A model that uses fewer steps per task is a model that decides sooner. It gathers less before committing, checks itself less often, and stops looking once it has something that fits. On a benchmark of multi-step tasks with a checkable answer, that is close to pure gain it is faster and it still lands.

The same disposition costs you on a forty-page document, where being right depends on going back and looking again at the paragraph that complicates the obvious answer.

That would also explain the reported feel. “More literal” is what decisiveness looks like from the outside: the model takes your sentence at face value rather than considering what else you might have meant, which is efficient when you meant what you said and irritating when you did not.

This is inference, not a published finding. Meta has not described the change this way and no benchmark isolates it. It is offered because it makes the three observations cohere and, more usefully, because it predicts where else to be careful anywhere being right depends on a second pass.

A model that checks three times before answering compared with one that commits after a single pass More steps ask look check again Slower. Catches the paragraph that changes the answer. ~20% fewer steps ask look the passes that did not happen Brisker. Also more literal, and less likely to look twice. Our interpretation, not a published finding
The step it skips is usually the redundant one. Occasionally it is the one that mattered.

Meta calls it efficiency

Around 20% fewer tool calls per task. Measured, published, and plainly good on work where the answer is checkable at the end.

Users call it literal

It takes the request as written. Excellent when you were precise, less forgiving when you expected it to read past your phrasing to your intent.

Testing calls it a regression

Long-document accuracy down four points. The one measurement of what committing early costs when the evidence is spread across many pages.

Use cases

Which of your work this helped, and which it did not

The figures only matter through the work they touch. Four broad kinds, with what the evidence actually supports saying about each.

Multi-step tasks that use tools

Clearly better

Work that chains actions together searching then comparing then drafting, or anything agent-shaped where the model has to decide what to do next several times before it is finished.

What supports this: both improved benchmarks are multi-step tool-use tests, so this is the one case where the gains were measured directly rather than inferred. It is also why Muse, which is built entirely out of such sequences, is the product most likely to feel the difference.

Careful reading of long documents

Measurably worse

Contracts, reports, research papers, long transcripts anything where the answer turns on a detail buried deep in the text rather than on the general gist of it.

What supports this: the independently measured drop from 83% to 79%. The practical response is not to avoid the task but to stop treating a confident summary as a reading, and to check the specific clause or figure yourself.

Ordinary chat, drafting and tidying

Probably unchanged

Rewriting a paragraph, answering a question, summarising something short. The bulk of what people actually do with an assistant on a Mac, and the part nobody publishes benchmarks about.

What supports this: no published figure, in either direction. Nothing measures everyday use, so the truthful answer is that the release probably did not change it much and that nobody has shown otherwise.

Anything built on the old behaviour

Worth re-checking

Prompts you tuned carefully, workflows that depended on the model's habits, instructions that worked because 1.2 interpreted them a particular way. A more literal model reads those differently.

What supports this: the reported shift in feel, plus the structural fact that a hosted model changes under you. If something you rely on started behaving oddly in early September, the model moving is a likelier explanation than anything you did.

Access

Where you meet it, and why you cannot pick a version

For nearly everyone, access to Muse Spark 1.3 is not a decision. It is what answers when you use Meta AI, and it arrived without asking.

Through the Meta AI app and the web

The Mac app is a window onto a hosted model, not something running on your machine, so you get whichever version Meta is serving. There is no version picker and no way to stay on 1.2.

Through Meta’s developer tools

Model identifiers, availability and any version pinning are matters for Meta’s own documentation, which this site deliberately does not copy. If your code needs to name a version, get the name from the source.

Not at all, if that is what you want

The only way to hold a model still is to hold the weights, which is a different product with a different set of costs see the open-weight page for what that actually involves.

Muse Spark replaced Llama across Meta AI in April 2026; 1.3 is a later release in that line rather than its beginning. The updates page tracks what has changed since, and our review covers how the Mac app feels to use on top of it.

Related:
FAQ

Muse Spark 1.3 questions

What is Muse Spark 1.3?

The model behind Meta AI, released on 2 September 2026. It has a one-million-token context window, accepts text, images and video, and improved on tool-use benchmarks while slipping on long-document accuracy. The four figures.

Is Muse Spark 1.3 better than 1.2?

Better at some things and measurably worse at one. Meta’s figures show gains on tool use Tau3-Bench Banking from 35% to 47% and Terminal-Bench 2.1 from 80% to 85% while independent testing found long-document accuracy fell from 83% to 79%. Whether it is an upgrade depends on which of those you rely on.

What does the one-million-token context window mean?

It can hold roughly a very large book’s worth of material in a single session, across text, images and video. Holding a document and reading it carefully are different things, though, which is the awkward part given the long-document result.

Why does it feel more literal than before?

Meta reports around 20% fewer tool calls per task, which means a model that commits to an answer sooner. The same tendency that makes it brisker on multi-step work is a plausible explanation for it double-checking less on long documents, though that connection is our reading rather than a published finding.

Which tasks got better?

Anything that chains several steps together with tools the benchmarks that improved are both multi-step tool-use tests, and agent-style work is where those gains should show up in practice.

Which tasks got worse?

Careful reading of long documents. If a detail forty pages in changes the answer, that is precisely the case where accuracy was measured to have fallen. What to do about it.

Can I choose which version I use?

Not in the Meta AI app. Model versions change server-side and the update arrives whether or not you wanted it, which is one of the structural differences between using a hosted service and running weights yourself.

Does the Mac app use Muse Spark 1.3?

Muse Spark is the model behind Meta AI, and 1.3 is the release that became the engine on 2 September 2026. The Mac app is a window onto that hosted model rather than something running on your machine.

Did Muse Spark replace Llama?

Yes. Muse Spark replaced Llama across Meta AI in April 2026, and 1.3 is a later release in that line rather than the first.

Should I trust these benchmark numbers?

Treat them as directional and note who produced each one. The tool-use gains and the reduced tool calls are Meta’s own figures; the long-document regression came from independent testing. A vendor publishing a result that flatters it is ordinary, which is part of why the independent one is worth its weight.

Two figures up · one down · one unmeasured

Ask which half of the trade you live in.

If your work chains steps together, 1.3 was a good month. If it turns on a clause buried in a long document, read that clause yourself.

Which work it helps How to read model figures
  • Released 2 September 2026
  • 1M-token context
  • Text, images and video in
  • No version picker

Independent guide not affiliated with Meta. Figures are attributed to Meta or to independent testing at the point of use. See the four in full.