Meta's Muse Glimmer ships under Apache 2.0 with no user cap and no commercial restrictions, a genuine break from the Llama community license. But the permissive terms apply to a 30B distillation of Muse Spark 1.2, not to the frontier model — and a model that runs on your own GPU moves the cost of inference from the vendor's balance sheet to your electricity bill. Both the license and the size are the same strategic decision.
On 10 August, Meta released Muse Glimmer, a 30-billion-parameter agentic model, with the weights published on Hugging Face under the Apache 2.0 license. Almost every account of it has led with the same sentence: Meta is back to open source. That is accurate, and it is the least useful accurate thing available. The more informative reading is that the license and the parameter count are not two separate decisions. They are one decision, and each explains the other.
Muse Glimmer is a distillation of Muse Spark 1.2 — a smaller model trained to reproduce the behaviour of a larger one — aimed squarely at agentic workloads: tool calling, local coding, retrieval-style research, and using a model to grade other models' output. Meta says it runs on a PC or Mac with a single consumer GPU. Quantized to roughly 4-bit precision, the language model comes in under 20GB. Secondary coverage cites a 24GB VRAM target, which is worth treating as a practical rule of thumb rather than a published requirement, since Meta's own materials describe the footprint rather than naming a card.
Meta published agentic benchmark scores alongside the release: 75.5 on MCP Atlas, 74.6 on DeepSearch QA, 47.6 on WildClawBench, 43.3 on GAIA2, and 23.5 on τ³-Banking, with additional evaluation on τ-Bench and SWE-Bench. Every one of those is Meta's own evaluation of Meta's own model. LM Studio, which distributes it, wrote that it is the strongest model of its size class the company has tested — that is an impression from a vendor whose business is local models, not an independent measurement, and it should be read as such.
The most informative thing in Meta's benchmark table is the spread, not the top score. A model that gets 75.5 on MCP Atlas and 23.5 on τ³-Banking is not uniformly capable at 'agentic tasks.' It is good at some and poor at others, and the label covers both.
Start with what genuinely changed. Meta's previous open-weight releases carried a bespoke community license that drew years of objection, most memorably for a clause requiring any company past roughly 700 million monthly active users to negotiate separate terms. It also carried an attached acceptable-use policy. Apache 2.0 carries none of that. It permits commercial use, modification, redistribution and sublicensing, with no revenue threshold, no user cap, and no obligation to publish your changes. It is recognised by the Open Source Initiative. For anyone who previously had to route a license question through legal before shipping a Meta model in a product, this is a material change rather than a cosmetic one.
Some coverage has put 'open source' in scare quotes, so it is worth resolving what the objection refers to. It is not a rider layered on top of Apache 2.0 — no additional acceptable-use restriction on the weights has been reported. The objection is about disclosure: Meta released the weights, not the training data or the training code. Muse Glimmer is therefore an open-weight model distributed under an open-source license, but it is not reproducible in the sense that someone could rebuild it from published materials. That is the standard complaint about weights-only releases, it applies here, and it is a complaint about what you can inspect rather than about what you are permitted to do.
Now the bounded part. Apache 2.0 applies to the 30B distillation. It does not apply to Muse Spark 1.2, the model the distillation came from. That is the entire trade, and it is a good one for Meta. Giving away a small model on permissive terms costs a frontier lab very little competitively — a 30B model is not going to displace anyone's flagship API — and buys a great deal: developers building on your architecture, your tool-calling conventions, your naming, and a default position in whatever the local-agent stack turns out to be. The generosity is real. It is also concentrated exactly where generosity is cheap.
Here is the part that mostly goes unwritten. When a model runs in a data centre and you call it over an API, the vendor buys the accelerator, pays for the power, absorbs the idle capacity, and charges you per token. When the same model runs on your laptop, all of those costs still exist — they have simply moved onto your hardware and your electricity bill. Nobody has made inference free. Someone else is paying for it, and that someone is you.
This is why 'runs on your device' is a business-model statement and not only a privacy statement. An always-on agent is the worst possible API customer: it polls, it retries, it runs in the background, and it generates enormous token volume for low marginal value per call. Those are precisely the workloads a vendor would rather not host. Meta's own framing — 'personal intelligence' — points directly at this. The model that watches your screen all day is a model you should be running yourself.
None of that makes the privacy argument a cover story. It is not. If a model runs entirely on your machine, your prompts do not leave it, there is no retention policy to read, no vendor-side logging to trust, and no subpoena target holding your history. That is a genuine and meaningful property, it is the main reason many people want local models at all, and it happens to align with the vendor's interest rather than conflicting with it. Both things are true at once, which is usually how durable strategies work.
Agentic benchmarks are less stable than the single numbers suggest. Unlike a knowledge test, an agentic evaluation measures a model inside a scaffold — the harness that gives it tools, decides how many steps it may take, handles retries, and formats results. Change the scaffold and the score moves, sometimes considerably, without the model changing at all. A score of 75.5 on MCP Atlas is a property of one model plus one harness on one date, not a capability constant you can carry into your own system.
The practical question is not whether a 30B local model beats a frontier API on capability. It does not, and Meta is not claiming it does. The question is where the trade is favourable. Local wins on latency for short calls, because you remove the network round trip entirely; on cost for high-volume, low-stakes, repetitive work, where per-token pricing accumulates faster than the amortised cost of a GPU you already own; on availability, because it works offline and does not break when a provider has an incident; and on privacy, unambiguously. It loses on hard reasoning, on very long context, and on anything where a wrong answer is expensive.
That maps to a fairly clear division of labour, and it is probably the shape the next year takes: a small local model handling the constant, cheap, private, latency-sensitive work, escalating to a frontier API when the task is genuinely hard. Meta would like to own the first half of that sentence. Apache 2.0 on a 30B distillation is a reasonable price to pay for it.
All benchmark figures cited are Meta's own published evaluations, as of the 10 August 2026 release. Hardware requirements reflect the quantized build and will vary by scaffold and context length.
Meta describes it as running on a PC or Mac with a single consumer GPU; quantized to roughly 4-bit precision the language model comes in under 20GB, and secondary coverage cites a 24GB VRAM target as a practical threshold. Treat that as a rule of thumb rather than a specification. What you lose at 4-bit is not uniform: quantization tends to be mildest on fluent text generation and harshest on long multi-step tool use, which is exactly what an agent spends its time doing. The published benchmark scores were not measured on the quantized build you will be running.
DrafterDaily covers AI releases by what the license, the parameter count and the benchmark methodology actually commit to — not by what the announcement implies.
More AI coverageClaude Fable 5.1 costs exactly what Fable 5 cost per token. The 25-to-45% saving Anthropic advertises comes from one repriced line item — cached input, now billed at 2.5% of list instead of 10%. That is a discount you only collect if you keep the agent running.
Anthropic was not breached. The malware was already on the customer's machine, and it took a session cookie rather than a password — which is why two-factor authentication did nothing and why server-side revocation is the only lever the vendor has.
OpenAI's incident report on the Hugging Face breach leads on a security failure. The remarkable part is a verbatim chain-of-thought in which a model identifies its action as unauthorised and proceeds anyway — and an escape route that was a package manager, not a superintelligence.