Meta’s 816B Muse Has a $20M API Line; 30B Glimmer Has None

Meta now offers two open-weight Muse models on different terms. Muse Spark Open, which Meta released on 15 October 2026, has 816 billion parameters and a licence line: a company that offers model access through an API and has more than $20 million in gross revenue in any 12-month period must sign a separate agreement with Meta first.

Muse Glimmer, which Meta released on 10 August 2026, is a 30-billion-parameter model under the Apache 2.0 licence that Meta says can run on a single consumer GPU. Meta says it trained Glimmer on the outputs of Muse Spark, the model whose open-weights version is Spark Open. Meta reports third-quarter results on 21 October 2026.

Business Pill · QUANTIZATION

A one-minute explainer of quantization. It teaches the general idea only and says nothing about any company in this story.

Two Models, Two Licences

Meta released the weights of Muse Spark Open on 15 October 2026: an 816B-parameter model with 42B active parameters and a 1M-token context window, which Meta calls “the most capable US open-weight model to date.” Meta says its licence allows many developers to use the model for free, with the $20 million API line above.

For Glimmer, Meta’s launch post of 10 August 2026 says Meta is “open sourcing the model weights under a permissive Apache 2.0 license.” The model card says all artifacts are released under Apache 2.0: full-precision weights, two 4-bit quantized versions, a drafter model for faster generation and the vision encoder.

In the Apache 2.0 licence and the usage policy linked from the model card, as we read them on 20 October 2026, we found no revenue threshold or API-hosting line. The usage policy lists prohibited uses, such as the unlicensed practice of a profession.

What Was Posted on X

TestingCatalog posted that a Confidential VMs feature is in the works on Muse web, citing a hidden setting that says it will first be limited to Meta AI subscribers. It is a sighting of an unreleased setting; TestingCatalog says Meta has previously teased the feature for before the end of 2026.

What Meta Says Glimmer Can Do

Meta’s launch post says Muse Glimmer is a 30-billion-parameter model “optimized for always-on local agent workflows” and is “small enough to run on a Mac or PC with a single consumer GPU.” The model card gives about 29.6 billion parameters, including a vision encoder of about 1.8 billion, and a context length of 131,072 tokens or more.

Meta writes: “At full precision, a 30-billion parameter model would require over 55 GB of memory.” The card says its 4-bit K-Quant-17GB version targets 24GB of graphics memory, and that on an Nvidia RTX 5090 the model went from 74.9 to 233.4 tokens per second with its drafter, a 3.1x speed-up, in Meta’s test.

On Meta’s own benchmark table, Glimmer scores 75.5 on MCP Atlas against 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B, and 51.2 on SWE-Bench Pro against 36.9 and 50.2. The same table shows Qwen3.6-27B ahead on SWE-Bench Verified (77.2 against 76.0) and OSWorld-Verified (75.6 against 65.9).

In its safety section, the card says Glimmer’s abilities are “approximately in line with other models in its size class, while showing strictly lower capabilities than larger open-weight models.”

Where the Two Models Meet

Meta says it trained Glimmer “on Muse Spark’s outputs using logit distillation, leveraging a similar data mix as the teacher.”

Spark Open is a related but separate release. Meta’s developer blog calls it “the open weights version of Muse Spark,” and its model card says this open-weights version has additional safety post-training, as we reported on 15 October. Meta released Glimmer’s weights on 10 August 2026, about two months before Spark Open’s.

Connecting the Coverage

On 15 October we reported Meta’s release of the Spark Open weights and the licence’s $20 million line for large API hosts.

On 19 October we reported that Meta’s Muse AI assistant passed 5 million downloads in three weeks, according to Sensor Tower data reported by Mashable, and that Amazon had begun blocking Muse from its retail site in September, as Bloomberg reported.

The same day, Meta said the first phase of Prometheus, its supercluster in New Albany, Ohio, was used to train its Muse foundational models, and that it is “working toward scaling it to comprise ~1 million GPUs.”

And on 20 October we reported that Section 5 of Meta’s Commercial Terms preview, effective 30 October 2026, reads: “You acknowledge and agree that AI features may make mistakes and may not always produce accurate, complete, or suitable results.”

Before Meta Reports Q3

Meta says it will release third-quarter 2026 results after the market closes on Wednesday, 21 October 2026. In its second-quarter release, Meta guided third-quarter revenue of $61 billion to $64 billion and 2026 capital expenditures, including principal payments on finance leases, of $130 billion to $145 billion.

Two questions are open before the release: whether Meta changes that capital expenditure range, and whether it gives any figure for revenue from the Muse models. Neither is known until Meta reports.

The Structural Read

As we read it, the licences draw the line in different places: Spark Open’s requires a separate agreement for API hosts with more than $20 million in gross revenue, while Glimmer’s Apache 2.0 terms set no such line in what we read.

As we read it, the hardware gap is part of the split: Meta says Spark Open runs on two M5 Ultra Mac Studios, and says a 4-bit Glimmer build targets 24GB of graphics memory.

As we read it, the two releases are connected by training, not only by name: Meta says Glimmer learned from Muse Spark’s outputs, and calls Spark Open the open-weights version of Muse Spark.

Three Implications

A LICENCE LINE AT THE TOP Spark Open’s licence requires API hosts with over $20 million in gross revenue in 12 months to sign a separate agreement with Meta.

A SINGLE-GPU TIER Meta says Glimmer is small enough to run on a Mac or PC with a single consumer GPU.

A RESULTS TEST Meta reports third-quarter results after the market closes on 21 October 2026.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework’s core rule reads: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize).”

As we read it, Meta’s two Muse licences can be read against that rule: the larger model carries a line for big API hosts, the smaller one ships under Apache 2.0. The framework was published in March 2026 and does not discuss Meta’s Muse models.

Update, 20 October: A Confidential Mode Spotted on Muse Web

TestingCatalog reported on X on 20 October that a Confidential VMs feature is “in the works on Muse web”, based on a hidden setting it found. It wrote that Meta “previously teased that this feature will go live before the end of 2026” and that Confidential VMs “will initially be available only to Meta AI subscribers.”

The setting’s text, as TestingCatalog shared it, reads: “Your conversations run in a secure, encrypted environment” and “Your Muse data won’t be used to improve AI at Meta.” TestingCatalog also wrote that voice mode, image and video generation, and wearables support won’t initially be available in the mode. This is a sighting of an unreleased setting, not a launch, and we have not seen Meta announce it.

What Is Not Established

Meta has not said how many companies have signed the separate Spark agreement, how many developers run Glimmer, or whether either model earns revenue. The benchmark and speed figures are Meta’s own, from its launch post and model card, and we have not run the model.

The Bottom Line

Meta’s Muse family now sits on two sets of terms: an 816B model whose licence requires large API hosts to sign a separate agreement, and a 30B model under Apache 2.0 that Meta says it trained on Muse Spark’s outputs and built to run on one consumer GPU.