How to install Muse Spark 1.3 now available on AI Gateway

What happened

Muse Spark 1.3 from Meta is now available on AI Gateway, in both the standard and contributor pricing tiers. If you already send traffic through AI Gateway, this is a model identifier change, not a new product to stand up. The model improves on prior Muse Spark models at agent work and coding. It has a 1M token context window and accepts text, image, and PDF input. On coding it takes fewer turns and writes less filler than the previous release.

To use Muse Spark 1.3, set model to meta/muse-spark-1.3. Keep that string exact. Do not invent a shorter alias, a date suffix, or a different vendor prefix. Confirm whether the call sits on the standard pricing tier or the contributor pricing tier, because both tiers carry this model. Then rerun the same agent and coding jobs you already trust so you can see the shorter loops and the quieter output without changing the rest of your stack.

How it works

Meta’s Muse Spark 1.3 is on AI Gateway today as a selectable model, not as a separate installer or a new gateway. The same model is offered on the standard pricing tier and on the contributor pricing tier, so access is not limited to one commercial path. What you get is an update over prior Muse Spark models for agent work and for coding, a 1M token context window, and input that can be text, image, or PDF. The coding change that shipped with this release is behavioral: fewer turns, and less filler than the previous release. The only identifier given for the new model is meta/muse-spark-1.3. No other version numbers, dates, or figures were provided, so do not invent them when you document the rollout. Treat the shipment as the model becoming available on AI Gateway under that identifier, on both named tiers, with the context window and input types listed above. If you already use a prior Muse Spark model on AI Gateway, this is the successor to point those calls at. If you do not yet use Muse Spark, this is the identifier to set when you add it. Nothing in the announcement describes a second SKU, a preview flag, or a regional split, so do not add those to your runbook.

Why it matters

Builders who already wire agents and coding jobs through AI Gateway should treat this as a quality and capacity change on a known path, not as a new integration surface. Agent work is called out as an area where Muse Spark 1.3 improves on prior Muse Spark models, which matters if your app loops a model through tools, files, and follow-up turns. Coding is the other named improvement: the model takes fewer turns and writes less filler than the previous release, so a coding agent should spend less time restating the task and more time producing the change. The 1M token context window is the capacity change. You can keep more of the repo, the ticket, the prior turns, and the attached files in one request, as long as you stay inside that window. Input is no longer a text-only assumption. You can send text, image, and PDF, which means a builder can attach a screenshot, a design export, or a PDF spec in the same call as the prompt. Pricing did not split the model into two different identifiers. Standard and contributor both get Muse Spark 1.3. The practical change is therefore: same gateway, same style of request, new model string, larger context, multimodal input, and tighter coding loops. Do not rewrite your product around a version number that was not given. Do not assume a new SDK. Change the model field, keep the rest of the request shape you already use on AI Gateway, and measure the agent and coding paths you care about.

First, find every place your app, script, or dashboard sets the model for AI Gateway. That is the only install or upgrade step required by this announcement. Second, set model to meta/muse-spark-1.3. Use that string as written, including the meta/ prefix and the 1.3 suffix. Third, confirm the project is on either the standard pricing tier or the contributor pricing tier. Both tiers include this model, so you do not install a different build per tier. You only confirm you are on one of those two. Fourth, if you are upgrading from a prior Muse Spark model, replace that model value with meta/muse-spark-1.3 and leave the rest of the request intact on the first pass. Do not change prompts, tools, or file handling until the new identifier is live. Fifth, send a text-only request you already know, then send the same job with image input, then with PDF input, so you prove all three input types on the new model. Sixth, run a coding task you already use as a baseline. Count the turns and read the output for filler. The previous release took more turns and wrote more filler; this release should take fewer turns and write less filler. Seventh, run an agent job that actually loops, not a single chat turn, because agent work is one of the named improvements over prior Muse Spark models. Eighth, keep the 1M token context window in view. If you previously truncated files or history to fit a smaller window, you can stop that truncation only after you confirm the new call is on Muse Spark 1.3 and that the combined text, image, and PDF payload still fits in 1M tokens. Ninth, save the old model string in your notes so you can roll back the identifier if a path is not ready. Rolling back is the same operation as upgrading: set model to the prior Muse Spark value you were using. Tenth, document for the rest of the team that the live identifier is meta/muse-spark-1.3, that AI Gateway is the place it is available, and that both pricing tiers carry it. There is no package to pin, no binary to download, and no extra version number to record beyond 1.3 as given.

Who is affected

The model string is the main compatibility surface. If you set anything other than meta/muse-spark-1.3, you are not on this release. Do not drop the vendor prefix. Do not write Muse Spark 1.3 as the model field. Do not invent a 1.3.0 or a date-stamped variant. Prior Muse Spark models are the thing this release improves on, so they are not the same model. Leaving an old identifier in one worker, one eval job, or one dashboard default means you will mix results and blame the new model for traffic it never saw. Standard and contributor are pricing tiers, not different model names. Do not create two identifiers to match the two tiers. The 1M token context window is a limit, not a promise that every payload is safe. Text plus image plus PDF in one request can still exceed 1M tokens if you attach large files and long history together. Count the whole request. Image and PDF input are supported, but only if you actually send those parts on the new model. An upgrade that changes the identifier and still strips non-text parts is a partial upgrade. Coding will not look better if you stop the agent after one turn. The improvement over the previous release is fewer turns and less filler, which you only see on a multi-turn coding job. Agent work is the other named gain; a single-shot prompt will not show it. Nothing in the announcement lists other version numbers, dates, or figures, so a runbook that adds them is wrong. If a tool, proxy, or allowlist still names only older Muse Spark models, add meta/muse-spark-1.3 before you send production traffic. If you cannot set the model field, you cannot install this release.

What to watch next

Watch the model field in production, not a changelog you do not have. Confirm live calls show meta/muse-spark-1.3 and that no worker has drifted back to a prior Muse Spark model. Watch coding jobs for turn count and filler against the previous release, because those are the two coding changes that were stated. Watch agent jobs for whether the loop is actually better, because agent work is the other stated improvement. Watch request size against the 1M token context window once you start sending longer history and mixed text, image, and PDF input. Watch that both the standard pricing tier and the contributor pricing tier keep serving the same identifier, since that is how availability was described. Watch any allowlist, proxy, or eval harness that might still block or reroute the new string. Do not watch for a version number that was not published. Do not wait on a date that was not given. If you need to undo the change, set model back to the prior Muse Spark identifier you replaced and keep AI Gateway as the path. If you keep the new identifier, keep using text, image, and PDF on purpose so the input support you installed is the input support you run.