Meta's newest open-weight model is the most capable you can self-host. The question is whether your hardware can fit 408 GB in memory.
816 billion parameters. Write that number down, then do the arithmetic before you get excited about running Meta’s new model on the Mac Studio under your desk.
Muse Spark Open is Meta’s entry into the open-weight race it had been ceding to Chinese labs for most of the past year. It takes text, images, video and audio, carries up to a million tokens of context, and — per Meta — trades blows with Qwen, GLM and Kimi, the three systems that have set the open-weight ceiling. That’s the headline, and it’s a real one. The frontier of what you can download and self-host just moved.
But a release like this is two stories. One is the capability. The other is what it costs to put the thing into production, and whether the “on-device, optimized for Apple silicon” line survives contact with a spec sheet. Let’s take them by the numbers.
816,000,000,000
The parameter count is the number everyone will quote and the one that tells you the least on its own. What matters is whether those 816 billion parameters are dense or a mixture-of-experts. Every open-weight model that has reached this scale and stayed usable — Qwen’s largest, Kimi’s K2, DeepSeek’s V3 lineage — is sparse MoE: a huge parameter store, a much smaller slice activated per token. If Muse Spark Open is dense at this size, inference economics would be brutal and the Apple-silicon claim would be a joke. So assume MoE, with an active parameter count a fraction of the headline.
Here’s the catch that MoE doesn’t solve: sparsity cuts compute, not memory. You still have to hold every expert in RAM or VRAM, because any token might route to any of them. The 816B is the number that sets your hardware floor, not the active count. Keep that straight and half the marketing evaporates.
4 modalities, 1 set of weights
Text, images, video, audio — in one model, not four stapled together behind a router. That’s the genuinely useful part, and it’s where open weights have lagged closed frontier systems hardest. Most self-hostable models that claim multimodality are strong on text, competent on images, and hand-wavy on video and audio. A single open-weight checkpoint that ingests all four with a long context is the thing a lot of teams have been waiting for, because it means you can keep regulated data — medical video, call-centre audio, internal documents — inside your own network instead of posting it to an API.
The honest caveat: “supports” a modality and “is good at” it are different claims, and vendors blur them on purpose. Video and audio understanding are where I’d spend my first afternoon of evals, not text. Text is table stakes at this point. If the video encoder chokes on anything longer than a short clip, or the audio path is really just a bolted-on transcriber, you’ll find out fast — and better now than in month six.
1,000,000 tokens
A million-token window is the spec that reads best on a slide and bites hardest in deployment. Context is not free RAM-wise — the KV cache grows with sequence length, and at seven figures of tokens it can rival the weights themselves for memory. A model that advertises 1M context will happily let you request it and then OOM your server at 300K because nobody budgeted for the cache.
Two things to verify before you trust the number. First, where was long-context performance actually measured — real retrieval across the full window, or a needle-in-a-haystack test that any model passes? Second, what’s the realistic cache ceiling on your hardware? Treat the million as a headroom figure, not a working default. The context you can afford to run continuously is the context you actually have.
~408 GB at 4-bit
This is the number Meta would rather you skip past. Run the weights math yourself:
Now hold that against Apple silicon. A maxed Mac Studio with the top M3 Ultra tops out around 512 GB of unified memory. So a heavily 4-bit-quantized checkpoint theoretically fits on exactly one, very expensive, configuration — with almost nothing left over for the KV cache that a long context demands. “Runs on Apple silicon” is true in the way “fits in your garage” is true of a bus.
What’s realistic on a laptop is a smaller quantized variant or a distilled version, served through mlx-lm, the Apple-native path that most of the usable local quants will land on within days of any release like this. The full 816B model is a data-center artifact. On NVIDIA you’re looking at multi-GPU tensor parallelism — eight cards is the usual starting point at this scale:
Note the max-model-len of 256K, not a million. That’s the trade you make to keep the cache on the cards. Push it higher only when you’ve measured the memory, not hoped for it.
3 rivals, and one claim to pressure-test
Meta positions Muse Spark Open against Qwen, GLM and Kimi — the open-weight leaders, all Chinese, all shipping aggressively. “Matches top Chinese systems” is the framing, and the per-benchmark numbers belong to Meta’s own release notes, which means you should read them the way you read any vendor’s self-selected scoreboard: as the best case, assembled by the people with the most to gain.
Vendor benchmark tables tell you a model can hit a score on a curated set. They don’t tell you how it behaves on your data, your languages, your weird edge cases. The useful move isn’t to argue about whether Meta’s claim is accurate — it’s to wait for the independent runs on neutral leaderboards and, more importantly, to run your own evals on the tasks you actually ship. A model that wins three public benchmarks and loses on your document pipeline is a loss. Trust the delta you measure, not the one you’re handed.
The strategic read matters more than the leaderboard anyway: Meta needed a credible open multimodal flagship to stop being the answer to a question nobody was asking in the open-weight world. On scale and modality coverage, it now has one. That changes the negotiating position for every team that was defaulting to Qwen because there was nothing comparable with a Western license.
The license is the real fine print
Expect the weights to land where Meta always puts them — a gated repository you accept a license to access, plus the usual request flow. And expect that license to be open-weight, not open-source, in the way Meta has done every time: you get the weights and broad permission to use them, inside terms that aren’t OSI-approved and come with Meta’s own acceptable-use and scale clauses. Read it before you build a product on it. “Open” in this market means “downloadable under conditions,” and the conditions are the part that bites.
Here’s where I land. Muse Spark Open is the most capable thing you can self-host today if you need all four modalities and a huge context in one model, and that’s genuinely worth something — especially for teams who can’t send data to an API and had no Western option at this tier. But the 816B headline is a lure. The decision isn’t made on parameter count; it’s made on whether you have 400-plus gigabytes of memory to spare, a KV-cache budget for the context you actually need, and a tolerance for a license that still won’t say “open source” out loud.
Download it. Quantize a variant you can afford. Run your own evals on video and audio, not Meta’s. Then decide. The specs are the easy part — they always are.