Meta's Muse Spark Push Leaves Llama's Roadmap in Doubt
Meta just shipped its best model yet, and you can't download it. Muse Spark 1.3 landed on September 3, 2026, and it's API-served only: no weights, no self-hosting, no fine-tuning it on your own cluster. If your org built its AI roadmap around Meta staying the open-weight champion, this is the week to reread that roadmap.
What happened with Meta Muse Spark and Llama
Muse Spark 1.3 is Meta's latest coding and agentic model, and on paper it's competitive. The "xhigh" configuration, the version developers can actually call today, scores 61 on Artificial Analysis's Intelligence Index, tying GPT-5.6 Sol max and Grok 4.6 high. A stronger "max" reasoning variant exists too, but it's still in internal safety testing and isn't broadly available, so the best number Meta is showing off isn't one you can build on yet.
Pricing didn't move from version 1.2: $1.25 per million input tokens and $4.25 per million output tokens on the standard tier. There's a cheaper Contributor tier at $0.10 / $0.20 per million tokens, but it requires letting Meta use your submissions for training, which is its own governance conversation for a regulated enterprise.
Here's the part that matters more than the benchmark: Muse Spark is closed, and it's the second Muse release in a row to be closed. Meta has said an open-weight Muse Spark release is coming, but hasn't committed to a version, a license, or a date. Through the first half of 2026, Meta shipped no new open-weight Llama model at all, and industry trackers now treat a near-term Llama 5 as a low-probability event rather than the default expectation. The frontier work, including Meta's in-development Avocado model, is staying proprietary. Llama looks increasingly like it's being kept alive as a mid-tier open option while the real capability investment goes into models you can only rent.
Why this is a leadership problem, not just an engineering one
A lot of enterprise AI deployments were built on the assumption that Llama would keep getting better and keep being downloadable. IBM watsonx.ai is a good example: it offers Llama 3.3 70B Instruct with IBM's governance layer, guardrails, audit logging, data residency, on top, which is exactly the kind of setup a bank or hospital picks specifically because the weights are open and the vendor risk feels lower than a closed API.
That calculation gets shakier when the vendor behind the weights stops shipping new ones. Llama 3.3 still works fine today, and nothing in Meta's announcement breaks existing deployments. But if you told a board or a procurement committee that open weights meant you weren't locked into one vendor's roadmap, you should go back and check whether that's still true. Right now, the newest, most capable thing Meta builds is API-only, priced by the token, and subject to whatever Meta decides Muse Spark 1.4 costs next quarter.
This isn't a Meta-specific problem either. It's a reminder that "open-weight" is a snapshot of a vendor's strategy at one point in time, not a permanent commitment. Any team that picked a model primarily because it was open needs a plan for what happens if the next version isn't.
What to do about it Monday morning
First, audit what actually depends on future open-weight Llama releases versus what's just running fine on Llama 3.3 today. Those are different risk levels. Second, get your cloud provider's current model catalog in front of whoever owns AI vendor strategy, so the open-weight options on the table are visible, not assumed.
# List foundation models available on IBM watsonx.ai to see current Llama coverage
ibmcloud login
ibmcloud wx model list --provider meta
This gives you a concrete list of what Meta models are actually deployable today, which is a better planning input than a roadmap promise. Third, if a project's business case leans on "we can always self-host if the vendor API gets expensive," pressure-test that against a model family, like Qwen or DeepSeek, that's still shipping open weights on a predictable cadence, so you have a real fallback and not just a hopeful one.
None of this means panic-migrating off Llama 3.3 today. It means treating "open-weight roadmap" as a line item you check every quarter, the same way you'd check a cloud vendor's pricing page, because the team that gets surprised by a strategy shift is the one that never asked.