Meta Releases Muse Spark Open, an 816B Open-Weight Model

Meta Returns to Open Weights, with 816B in Tow

On October 15, Meta Superintelligence Labs released the weights for Muse Spark Open.¹ It’s a Mixture of Experts (MoE) model with 816B total parameters, where only 8 of its 256 experts activate per token, so the active parameter count actually used in computation is 42B. The model card states it has a 1 million token context window, takes text, image, video, and audio as input, outputs text, and was trained on more than 100 languages.² Meta describes this as the most capable open-weight model to come out of the US. Open weight means you can download the trained weight files and run them yourself—unlike open source, which also opens up the training data and code.

Given Meta’s moves this year, this release is a bit of a surprise. When Meta Superintelligence Labs’ first model, Muse Spark, launched on April 8, the weights weren’t open. It was the first closed model from a company that had become synonymous with open models through the Llama series, and at the time Meta only said it hoped to open the next version.³ On August 5, Muse Spark 1.2 and the terminal coding agent Muse Code also shipped only as an API and a product.⁴ Then, half a year later, Meta put out an open-weight version of 1.2.

Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to DateMuse Spark Open is the most capable US open-weight model to date, and brings leading agentic capabilities to your desktop on Apple M5 Ultra Mac Studios.research.meta.ai

Two things caught my attention in this announcement. One is that the machine Meta pointed to when saying you can “run it at home” is Apple’s Mac Studio M5 Ultra. The other is that we might be about to see a replay of the scene from Llama 2 back in 2023. This issue walks through both in turn, then moves on to the memory story that follows from them.

It Runs on a Desk—If You Have Two Mac Studios

Here’s how Meta’s blog post describes it. The M5 Ultra Mac Studio lets its GPU tap up to 512GB of unified memory, with 1.2TB/s of memory bandwidth. Unified memory means the CPU and GPU share the same memory pool, so you can load an entire large model regardless of the usual VRAM ceiling on a graphics card. Chain two of these machines together and they behave like a single supercomputer with 1TB of unified memory—big enough to hold the whole model.¹ Meta says it optimized with Apple’s MLX framework to run Muse Code on this setup, and completed every task inside the machine itself: patching and rebuilding the iOS code for Ice Cubes, an open-source Mastodon client, and using DaVinci Resolve to inspect video frame-by-frame and crop it vertically.

One machine isn’t enough. The baseline Meta specifies is two 512GB Mac Studios, 1TB combined. Apple’s machine learning research team posted the numbers the same day. Even compressed to 4-bit NVFP4, the weights run about 493GB; the original BF16 version exceeds 1.6TB—neither fits on a single machine. Chain two together with a Thunderbolt 5 cable and let them exchange memory directly via the RDMA driver built into macOS 26.2, and generation speed reaches 15.7 tokens/sec (pipeline parallelism) or 25.0 tokens/sec (tensor parallelism). Chain four together and it climbs to 36.7 tokens/sec. Time to first token is 8.2 seconds with two machines.⁵ Meta’s model card is more conservative, stating a minimum of 700GB of memory for NVFP4 and 2TB for BF16.²

That’s not to say squeezing it onto one machine is impossible. The community-built 2-bit GGUF version weighs in at 304GB, which does fit on a single 512GB Mac Studio, and there are reports of a Hugging Face team running it on an M3 Ultra 512GB and getting roughly 18 tokens/sec. But 2-bit quantization means sacrificing quality, and if you use the full 1-million-token context, the KV cache alone takes up about 266GB—combined with the weights, that exceeds what a single machine can hold.⁶ “Runs on one machine” is true for a demo, but comes with conditions attached for real work.

NVIDIA’s DGX Spark isn’t even close. Its unified memory is 128GB, and chaining two together only gets you to 256GB.⁷ That’s despite the machine launching at $3,999 last October and jumping to $4,699 this February, officially due to memory shortages. A machine once billed as a personal AI supercomputer turns out to hold just 1/8 of the 1TB Meta is talking about for this model. If NVIDIA wants to compete in this weight class on a desk, it has to move up to the GB300-equipped DGX Station—748GB of memory, priced at $92,887 through Supermicro.⁸

The Mac Studio side isn’t cheap either. The M5 Ultra configuration starts at $5,499 for 96GB, and jumps to $12,299 for 256GB.⁸⁹ The 512GB configuration ships separately in late October, and no price had been announced as of the August unveiling. And you need two of those machines. The previous generation’s M3 Ultra 512GB configuration was pulled from sale this past March, officially due to a memory chip shortage, and there were reports of used units in Australia selling for more than $60,000.¹⁰ So yes, “it runs on a desk” is accurate. The condition is that the desk holds two machines, each costing well over $10,000.

Memory Bandwidth Sets the Ceiling on Token Speed

Let’s work through why Meta wrote that “1.2TB/s is enough to read what’s needed for each token.” An MoE model only reads its active parameters to generate a single token. Storing 42B parameters at 4-bit precision takes about 21GB per token. At 1.2TB/s of memory bandwidth, you can read that 21GB up to 57 times per second—ignoring every other cost, that puts the ceiling at 57 tokens/sec. The 25 tokens/sec Apple measured on two machines is roughly half that ceiling, with the rest spent on KV cache reads, routing, and communication between the two machines. Run the same math on the DGX Spark, with its 273GB/s of bandwidth, and the ceiling is 13 tokens/sec—that’s the speed even if the weights do fit.

This math reveals something important. Because only 42B out of 816B parameters are active, the compute load resembles that of a 42B model—but all 816B still have to sit in memory. MoE trades away compute in exchange for demanding memory. That’s why running a model in this weight class at home becomes a question of memory capacity and bandwidth before it’s ever a question of GPU power. It’s why Apple walked into this space leading with 512GB of unified memory, and why memory shortages come back up later in this piece.

The question of why anyone needs those specs

This scene reminds me of smartphone cameras. For a while, people kept asking why anyone would need that kind of camera on a high-end phone. Look around now — everyone’s carrying one.

The iPhone Duo follows the same pattern. At the September announcement, the $1,999 price tag drew plenty of complaints about being too expensive. Since rumors had pegged it at up to $3,000, some reactions were actually relief that it was cheaper than expected — but it’s still the most expensive iPhone ever made.¹¹ In Korea it starts at ₩3,290,000 (~$2,370), with preorders opening today, October 16, and launch set for the 23rd.¹² In the US, when preorders opened this morning, launch-day delivery slots sold out within minutes, and the delivery window slipped from late November to December.¹³ Granted, there’s also speculation that initial production volumes are lower than usual — meaning the sellout isn’t proof of demand alone. Still, the fact remains: people lined up for something everyone called overpriced. I don’t think this supply shortage will be a one-time thing.

So one question follows. Is an era really coming where every household needs its own supercomputer? I’m still keeping a question mark on that one. But I do remember how the question “why would an individual ever need those specs” turned out last time.

The Great Build-Out Era Llama 2 Started

Meta’s open-model policy has had an outsized impact on LLM development. The first Llama, released in February 2023, was open only to researchers; everything changed that July when Llama 2 allowed commercial use too.¹⁴ Many of the early models that followed borrowed everything from Llama — the model card, the architecture, even the inference code. China’s 01.AI turned out to have simply renamed two tensors in the Llama architecture for its Yi-34B, and the company admitted the mistake and reverted the names,¹⁵ while Korea’s Upstage wrote right in the model card for SOLAR-10.7B that it was built on the Llama 2 architecture.¹⁶ Europe was no different. Using the same framework meant these models dropped straight into tools like llama.cpp, and the next generation of models then emerged on top of those same tools. I call that period the great build-out era of LLMs.

I think this release could bring that scene back. On the inference-engine side, SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth were all on board from day one, with optimizations tuned for AMD, Apple, Intel, and Nvidia chips. The weights are available not just on Hugging Face but on Amazon SageMaker, Google Cloud’s Model Garden, and Microsoft Foundry, and Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together AI all opened API access on day one.¹ Back in 2023, the community built those connections after the fact; this time Meta laid them all down before the release. Whatever model comes next starts out with this tooling compatibility built in as a baseline.

What “the best from America” is really pointing at

Why Meta bothered to attach “made in America” in front of “the best open-weight model” becomes clear once you look at the comparison table in the model card. Every model in the comparison is open-weight, and the ones in a similar weight class are Alibaba’s Qwen3.8 (2.4T parameters, 95B active), Zhipu’s GLM 5.3, and Moonshot’s Kimi K3.² Kimi K3 is a 2.8T-parameter model (total) released in July.¹⁷ Let me pull a few rows from the table Meta published. Keep in mind these are benchmarks Meta chose, run by Meta.

Benchmark	Muse Spark Open	Qwen3.8	GLM 5.3	Kimi K3
AA Intelligence Index	39	40	45	44
JobBench	61.8	52.7	61.4	54.3
MCP-Atlas (tool use)	89.2	84.5	84.2	82.3
Terminal-Bench 2.1	76.4	82.0	83.9	85.0
GPQA Diamond	89.0	93.5	91.7	93.5

It leads on tool-use and job-style benchmarks, but trails all three Chinese models on the composite index and on terminal and science problems. So “the best from America” isn’t wrong, exactly — it just means the comparison set was Chinese open models from the start. I think that’s also why Meta came back to open weights. There’s no way Meta doesn’t know which country’s models the developers downloading open models have been building on for the past year. The full evaluation report hasn’t come out yet.¹

The conditions attached to the term “open weight”

There’s also something different from the Llama 2 era worth noting. The license. Meta released this under the Muse Model License, and it comes with a few conditions attached. You have to display a “Built with Muse Spark Open” notice on the product screen or in documentation, and if your business offers this model to others via API and earns over $20 million in trailing-12-month revenue, you need a separate agreement with Meta. Internal use, research, and embedding it in your own product are exempt from that requirement. Any model trained, fine-tuned, or distilled from this model or its outputs counts as a derivative.² This is open weight, not open source. That’s nothing new — the OSI never recognized Llama as open source either,¹⁴ but if you build a product without knowing these conditions, they’ll catch up with you later.

The weight class has shifted — 1T is now the baseline

816B looks like a massive number, but in this year’s open-weight landscape, it isn’t. Kimi K3 sits at 2.8T, Qwen3.8 at 2.4T. Total parameter counts around 1T have become the expected norm even for open-weight models, and the default design now activates only a few dozen B of that. Muse Spark Open activates 42B; Qwen3.8 activates 95B.²¹⁷

Bringing back the earlier math, what this design demands from hardware isn’t compute — it’s memory. With 42B active, the compute load is mid-size-model territory, but all 816B of weights still have to sit in memory, and if you’re running a 1,000,000-token context, the KV cache adds several hundred more GB on top. The memory a single model demands has climbed from tens of GB to the terabyte range in just two years. As more people try to run this locally, the picture becomes one where individuals and small organizations end up buying two 512GB machines apiece. I see this release as yet another push further into memory scarcity.

Who’s Supplying the Memory

Memory is now the hardest component to get your hands on. TrendForce’s July survey found that third-quarter contract prices for commodity DRAM would rise 13–18% over the previous quarter. The increase is smaller than before not because supply loosened up, but because PC and smartphone customers have hit a price ceiling they simply can’t absorb anymore.¹⁸ You can see the fingerprints of this on actual products. Apple pulled the 512GB Mac Studio in March because of a memory chip shortage and has only now brought it back, while Nvidia raised the price of the DGX Spark for the same reason.⁷¹⁰

The number that hits closest to home for Korean readers came out on October 8th. Samsung Electronics’ preliminary operating profit for the third quarter was ₩107.4 trillion (~$77.4 billion). That’s the first time quarterly operating profit has topped ₩100 trillion (~$72 billion), on revenue of ₩195 trillion (~$140.6 billion) and an operating margin of 55%. The jump comes from rising HBM4 sales for Nvidia’s next-generation GPUs plus higher commodity DRAM prices. But within the same company, you can see the opposite effect playing out. The device division, which handles smartphones and home appliances, is seeing its losses widen because memory prices have gone up, and analysts expect a loss of roughly ₩2 trillion (~$1.4 billion).¹⁹ The side selling memory is posting record profits; the side buying memory to make products is bleeding money, even though they’re the same company. It’s the same story as the iPhone duo selling out. While data centers hoover up all the HBM (high-bandwidth memory) they can get, and supply stays tight, nobody’s left to ask “why do we even need that spec” — which means there’s no room left for prices to come back down.

Oswarld’s Lens

On my desk, an M2 Max 64GB Mac Studio runs as my inference server. I load it with a 30B-class model and work from my MacBook connected to it. This machine can’t come close to running Muse Spark Open. And yet a good chunk of my work runs just fine at this scale. As I keep saying, not everyone and not every task needs superintelligence. Different jobs and environments call for different models. I hand overall planning and orchestration to a frontier model, while offloading the long, repetitive work to cheaper open models.

That’s why another figure Apple showed in the same piece caught my eye. A 27B dense model running on a single M5 Ultra hits 19.7 tokens per second using 68GB of memory.⁵ This is still the weight class that runs at home. The picture this announcement paints looks more like a split between home and organization: tens of gigabytes of memory and a 30B-class model at home, versus 1TB of memory and a 1T-class model for organizations.

So before I think about “a supercomputer in every home,” I look at something else first: the air-gapped network environments of public institutions. When I design a project for a closed network, I set the specs by estimating what open-weight models will look like 6-12 months out, on the assumption that models only ever get better over time. This announcement moves that yardstick up a notch. For organizations that can’t send data outside their walls, “a near-frontier model running on my own machine” has, for the first time, become an option they can put on a realistic budget estimate. Two machines priced above $10,000 are a lot of money for an individual, but they’re a small line item next to a $90,000 DGX Station or a GPU server budget.

I still don’t know whether I’ll bring one home. But smartphones already showed us that the question “why would an individual ever need that kind of spec” has a shorter shelf life than we’d expect.

Closing

Meta is back with open weights, and the new baseline is two Mac Studios and 1TB of memory. Memory supply simply isn’t keeping pace with how fast that baseline is becoming the norm.

---

💬 Reader, what machine are you running which models on, at home or at the office? Or are you still in the “why would anyone need that much spec” camp? Let me know in the comments.

📨 If you know someone weighing whether to try running models locally, pass this piece along to them.

References & Further Reading

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

• Open weight: A release method where the trained model’s weight files are published for download, execution, and fine-tuning. This does not mean the training data or training code is also open, and it’s distinct from an OSI-compliant open-source license.
• MoE (Mixture of Experts): An architecture where, for each token, only some of many expert sub-networks are selected for computation. All parameters still have to be loaded into memory, but the computation cost scales with active parameters, not total parameters.
• Active parameters: The number of parameters actually used in computation when an MoE model processes a single token. For Muse Spark Open, that’s 42B out of 816B total.
• Unified memory: An architecture where the CPU and GPU share the same memory pool. This is why a GPU can load an entire large model despite the usual VRAM limits of a dedicated graphics card.
• Memory bandwidth: The amount of data that can be read from memory per second. Because the generation phase requires re-reading the active weights for every single token, this figure sets the upper bound on tokens generated per second.
• NVFP4: A 4-bit floating-point format defined by Nvidia. It shrinks weight storage to roughly a quarter of what BF16 requires.
• KV cache: Memory that stores the keys and values of tokens already processed. It takes up memory separately from the weights, and grows as the context gets longer.
• Tensor parallelism / pipeline parallelism: Two ways of splitting a model across multiple machines. Pipeline parallelism splits the model front-to-back by layer; tensor parallelism splits each layer widthwise, so every machine handles a slice of every layer.
• RDMA (Remote Direct Memory Access): A method that lets one machine read and write another machine’s memory directly, without going through the CPU. Supported over Thunderbolt 5 starting with macOS 26.2.
• HBM (High Bandwidth Memory): Stacked memory used in data-center AI accelerators. As memory manufacturers redirect production capacity toward HBM, the supply of ordinary DRAM has been shrinking.

Footnotes

1. Meta Superintelligence Labs, Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date, 2026-10-15. Two-Mac-Studio configuration, inference engine and deployment path, full evaluation report not disclosed. ↩ ↩² ↩³ ↩⁴

2. Meta, Muse-Spark-Open-816B-A42B model card, 2026-10. 256 experts with 8 active, input modality, 100+ languages, NVFP4 minimum 700GB / BF16 2TB, Muse Model License terms (attribution requirement, $20 million MaaS revenue threshold, derivative-work definition), benchmark comparison table (Qwen3.8-2.4T-A95B, GLM 5.3, Kimi K3, etc.). ↩ ↩² ↩³ ↩⁴ ↩⁵

3. The Next Web, Meta’s Muse Spark is MSL’s first model, 2026-04-08. The first Muse Spark’s unannounced launch, intent to release future versions. ↩

4. Meta Superintelligence Labs, Introducing Muse Code and Muse Spark 1.2, 2026-08-05. Muse Code and the Meta Model API path, no mention of weight release. ↩

5. Apple Machine Learning Research, Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio, 2026-10-15. NVFP4 ~493GB / BF16 1.6TB+, 15.74/24.99 tok/s on 2 units, 36.73 tok/s on 4 units, 8.2-second time-to-first-token, macOS 26.2 RDMA, 27B dense model at 19.69 tok/s / 68GB on a single unit. ↩ ↩²

6. ModelFit, Meta Muse Spark Open on a Mac: 304GB 2-bit Build Needs a 512GB Mac Studio, 2026-10-16. ggml-org GGUF sizes (Q2_K 304GB, Q4_K_S 470GB, Q8_0 872GB), Hugging Face team’s ~18 tok/s on an M3 Ultra 512GB, ~266GB KV cache for 1M context. Secondary source. ↩

7. Toolhalla, NVIDIA DGX Spark Complete Guide 2026; Implicator, Nvidia ships DGX Spark at $4K. 128GB unified memory, 273GB/s bandwidth, 256GB when linking 2 units, launch price $3,999 on 2025-10-15, $4,699 in 2026-02 citing memory shortages. ↩ ↩²

8. TechRadar, Nvidia GB300 DGX Station desktop PC goes on sale for almost $91,100, 2026-09-01. Supermicro selling price $92,887.47, 748GB (496GB CPU + 252GB HBM3e), Mac Studio M5 Ultra 256GB at $12,299. ↩ ↩²

9. Thurrott, Apple’s Mac Studio is Being Refreshed With M5 Ultra and M5 Max Chips, 2026-08-25. Starting price $5,499 (96GB / 1TB), 512GB configuration launching late October with price undisclosed. ↩

10. MacRumors, Mac Studio With M5 Ultra Chip and 512GB of RAM Launching in October, 2026-08-25. 512GB configuration sales halted in 2026-03 due to memory chip shortages. Secondhand prices exceeding $60,000 reported by Man of Many (Australia), Mac Studio M5 Max M5 Ultra. ↩ ↩²

11. PhoneArena, People thought $3,000 for a foldable was too much and Apple listened, 2026-09. Starting price $1,999 versus rumored $3,000. ↩

12. WikiTree, Apple unveils foldable ‘iPhone Duo’ — Korean pricing ‘starts at ₩3.29 million (~$2,370)’, 2026-09. Korean pricing, pre-orders October 16, launch October 23. ↩

13. AppleInsider, iPhone Duo sells out moments after pre-orders begin, 2026-10-16; Macworld, iPhone Duo pre-order live blog, 2026-10-16. Shipping dates of November 16–23 set immediately after pre-orders opened, later pushed to December. The observation that initial stock was lower than usual comes from AppleInsider. ↩

14. Wikipedia, Llama (language model). First released 2023-02-24 under a non-commercial research license, commercial use permitted from Llama 2 in 2023-07, not recognized as open source by the OSI. ↩ ↩²

15. South China Morning Post, Chinese tech unicorn 01.AI admits ‘oversight’ in changing name of AI model built on Meta Platforms’ Llama system, 2023-11. Original discussion at Hugging Face 01-ai/Yi-34B discussion. ↩

16. Upstage, SOLAR-10.7B-v1.0 model card, 2023-12. Explicitly states “Built on the Llama2 architecture.” ↩

17. Sébastien Dubois, Kimi K3, Qwen 3.8, and a Week That Reset the Open-Weight Frontier, 2026-07. Kimi K3 total 2.8T released July 16, Qwen 3.8 announced July 19. Secondary source; the 95B active-parameter figure for Qwen3.8 follows Meta’s model card notation. ↩ ↩²

18. TrendForce, AI Server Demand Continues to Support Memory Prices in 3Q26, but Gains Moderate as Consumer Demand Weakens, 2026-07-03. General-purpose DRAM contract prices projected to rise 13–18% quarter-on-quarter in 3Q26, limits to consumer-market price absorption. ↩

19. MTN (Money Today Broadcasting), Samsung Electronics opens the era of ₩100 trillion quarterly operating profit (comprehensive), 2026-10-08. Q3 preliminary revenue ₩195 trillion (~$140.4 billion), operating profit ₩107.4 trillion (~$77.3 billion), margin 55%, driven by HBM4 and general-purpose DRAM price effects; DX division projected to post a loss of roughly ₩2 trillion (~$1.4 billion). ↩