Muse Spark Open explained: how an 816B model fits on two Mac Studios

Yesterday Meta published the full weights of an 816-billion-parameter model, and Apple's chief executive John Ternus replied on X that it was "amazing" to run a model of that scale "right from your desk". A year ago those two sentences would have been a contradiction: models that size lived in data centres. What changed is less about the model's intelligence, which by Meta's own numbers trails three Chinese open models, and more about three pieces of engineering finally lining up so that a frontier-scale network fits in a terabyte of memory that you can buy.

The plain-English version

Think of a large AI model as an enormous reference library. Every time the model writes one word, it has to walk through the shelves and consult the books it needs. For years the biggest models had so many books that the library would not fit in any building you could put on your desk, so they had to be rented, shelf by shelf, from cloud companies.

Meta's new model, called Muse Spark Open, has 816 billion "parameters", which are the numbers the model learned during training. Stored at normal precision that is about 1.6 terabytes, roughly the contents of a few hundred laptops' worth of RAM. On 15 October Meta released those numbers for anyone to download and, together with Apple, showed the model running on two Mac Studio desktop computers plugged into each other with a cable.

Two tricks make that possible. The first is that this library is organised into 256 specialist sections per floor, and for any given word the model only visits eight of them. Meta's model card says only 42 billion of the 816 billion parameters are used for each word. The whole library still has to exist, but the librarian only walks a short route each time, which is why it can produce text at a readable pace.

The second trick is compression. Meta ships a version where each number is stored in 4 bits instead of 16, which Apple's engineers measured at about 493 gigabytes. Apple's newest Mac Studio can hold 512 gigabytes of memory that the graphics chip can use directly. Two of them, joined by a Thunderbolt 5 cable, add up to a terabyte, which Meta describes as making the pair "behave like one supercomputer".

The catch is price and performance. The 256GB Mac Studio starts at $10,799, the 512GB version is not out until later this month, and Apple's own tests show the two-machine setup produces about 25 words-worth of tokens a second, with a long document taking the better part of a minute before the first word appears. And "open" has conditions: businesses that earn more than $20 million a year selling access to the model must sign a separate deal with Meta.

If you take one thing away: the newsworthy part is not that Meta built the smartest open model, because it did not. It is that a model of this size can now run entirely on hardware a small company or a well-funded hobbyist owns, with no cloud account and no data leaving the room.

How it actually works

Running a large language model has two distinct costs, and the two Mac Studios story only makes sense once you separate them.

Cost one: holding the model. Every parameter must sit in memory that the processor can reach quickly. For Muse Spark Open that is 816 billion parameters. Meta's model card lists requirements of roughly 2TB for the BF16 version and at least 700GB for the NVFP4 version, where the 700GB figure includes working space beyond the weights themselves. Apple measured the NVFP4 weights alone at 493GB; Baseten, one of the launch hosting partners, quotes about 540GB for its serving setup. The difference reflects what each counts, but the point is the same: it does not fit on one 512GB machine, and it does fit on two.

Cost two: reading the model per token. When a model generates text it produces one token at a time, and for each token it reads the parameters it needs. On modern chips this step is limited not by arithmetic but by memory bandwidth, meaning how many bytes per second the chip can pull from RAM. Apple's post states this plainly: generation speed is "memory-bandwidth-bound", and its decode measurements scale almost exactly with the M5 Ultra's 1.2TB/s versus the M3 Ultra's 819GB/s.

This is where the mixture-of-experts design matters. In a dense model, every parameter is read for every token. In Muse Spark Open, each of the 62 layers contains 256 expert sub-networks and a router that picks 8 of them per token. Only those 8 experts, plus the shared attention layers, are read. The model card puts the active count at 42B parameters. At 4 bits per parameter that is about 21GB per token, which 1.2TB/s of bandwidth could in principle serve more than 50 times a second if nothing else got in the way. The observed 25 tokens a second on two machines is lower because of communication and overhead, but it is in the right ballpark, and that is the whole trick: you pay for 816B in memory capacity and for 42B in memory bandwidth.

Joining two machines. Here Apple contributed the new piece. macOS 26.2 adds a low-latency RDMA driver over Thunderbolt 5, and MLX version 0.30.1 adds a communication library called jaccl built on top of it. Apple's post describes two ways to split the model:

Apple's numbers, all for the NVFP4 model with an 8,192-token prompt and 128 generated tokens, show the difference:

Source: Apple Machine Learning Research. These are the 8,192-token columns of Apple's tables; Apple also reports 65,536-token runs, discussed below.

Tensor parallelism only works because the Thunderbolt link is fast enough that the per-layer exchange does not swamp the gain. Apple notes that at a 65,536-token prompt, two-machine tensor parallelism actually processed the prompt more slowly than pipeline mode, because of communication overhead, and that doubling the cables between four machines cut time-to-first-token from about 50.5 seconds to 40.6 seconds. Decoding, the part the user watches, was not affected by that overhead.

What the model is for. Meta pitches it at coding and agentic work. Its demos on the Mac setup used its own Muse Code agent to ship changes to Ice Cubes, an open-source Mastodon client, and to drive DaVinci Resolve to find and crop video highlights. The model takes text, images, video and audio in and produces text out, has a 1-million-token context window, and exposes four reasoning-effort levels, with "xhigh" as the default and the setting used for all its published benchmarks.

The deep dive

Architecture

The model card specifies a 62-layer decoder with hidden dimension 8,192, a 202,048-token vocabulary with untied input and output embeddings of roughly 1.66B parameters each, and grouped-query attention with 128 query heads, 16 key-value heads and head dimension 64. Attention is gated, with an output gate and output norm, and uses QK-norm. The MoE block has 256 experts with top-8 routing. Multimodal input goes through a separate Perception Encoder of about 1.8B parameters, 50 layers, 1,536 hidden width, 14×14 patches and 4,096 output dimension, with full attention only on every fourth layer and the last. Baseten gives the active count as 42.7B; Meta rounds to 42B.

Two design choices, both described in Baseten's launch note rather than Meta's announcement, explain how the model keeps memory and bandwidth in check.

Compressed MoE. A shared "pooler" projects each token's 8,192-wide residual vector down before it enters the experts, and a shared "unpooler" projects the combined expert output back up. The router still reads the full-width vector to choose experts. Baseten says this halves each expert's parameter count, saving roughly 800B parameters against an uncompressed design with the same expert count, and cuts inter-GPU traffic in the MoE layer because the vectors exchanged between devices are narrower. This is the component Apple's post is alluding to when it mentions that "separate normalization of experts' outputs" adds prefill communication in tensor-parallel mode.

Hybrid local-global attention. Of the 62 layers, 46 use a 2,048-token sliding window and 16 attend over the full context. The KV cache per sequence in BF16 is therefore

where $L$ is the context length, the leading factors are keys-and-values, KV heads, head dimension and bytes per element. At $L = 131{,}072$ that gives about 8.4GiB, versus about 31GiB if all 62 layers were global, the 3.7× saving Baseten reports. At the full 1M context the cache is still about 64GiB per sequence, a figure Kingy AI derives from the same formula. Note that the Mac Studio budget has no room for that: with roughly 500GB of weights in 1TB, a single 1M-token session is theoretically possible but leaves little for activations, and Apple's benchmarks stop at 65,536 tokens.

Quantisation and serving

Meta ships BF16 and NVFP4 checkpoints. NVFP4 is Nvidia's 4-bit floating-point format with per-block scaling, and the GPU serving path uses it natively: Nvidia's Dynamo recipe runs NVFP4 weights with an FP8 KV cache on four GB200 GPUs with 4-way tensor and expert parallelism at 1M context, and Baseten says the model fits on a single 4×GB300 node. Meta also contributed auto-tunable Helion kernels to SGLang, claiming 1.03-3.95× kernel-level speedups on Nvidia and AMD GPUs and 3% end-to-end on AMD MI350. Supported engines at launch are SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth.

On the Mac path, Apple's requirements are macOS 26.2 or later, MLX 0.30.1 or later and the NVFP4 checkpoint, which Kingy AI puts at about 537GB on disk. Apple's four-machine setup is a Thunderbolt mesh in which each pair of machines is linked peer-to-peer by two cables. The four-machine configuration with 256GB each works today; the two-machine configuration needs the 512GB Mac Studio that ships in late October. Apple measured peak power of about 300W per M5 Ultra, though only for its Qwen and Mistral runs, not for Muse Spark Open.

Benchmarks

All numbers below are from Meta's model card, measured by Meta at xhigh effort unless noted, and the comparison set is open-weight models only. Meta's table excludes the closed frontier models.

Source: Meta model card on Hugging Face, vendor-reported. For space, this table drops two columns that Meta's includes: Reflection AI's Beam, a US model whose weights are not yet public, and Mistral Large 4. The Artificial Analysis Intelligence Index row matches the independent listing Artificial Analysis has since published, which gives Spark Open 39 on index v4.3.2 at a cost of $0.84 per task. VentureBeat notes the same site scores MiMo-V2.6-Pro at 46, above every model in Meta's table.

The pattern is consistent. Against the other US entries, Inkling and Nemotron 3 Ultra, the lead is wide everywhere. Against the Chinese models the model wins on tool-use agent benchmarks (MCP-Atlas, JobBench) and on knowledge without tools (HLE-Diamond), splits the vision suite with Kimi K3 (ahead on ZeroBench, SimpleVQA and IntPhys2, behind on BabyVision, WorldVQA and VideoMMEv2), and loses on coding (Terminal-Bench, DeepSWE), on GDPVal-AA and on the composite index. The StrongREJECT number is the outlier in the other direction: 0.5 versus 60-plus for everyone else means Meta's safety post-training makes it refuse adversarial jailbreak prompts far more often than its peers.

Two harness caveats matter for the coding rows. Baseten notes that in agentic coding comparisons each model ran with its own lab's agent, so the numbers mix model and tooling. And the model card itself labels DeepSWE and SWE-Atlas as "preferred harness" results.

Safety evaluations

Meta says the model was evaluated under its Superintelligence Scaling Framework and met release thresholds; the full evaluation report is "available soon". The card reports 99.2% refusal of cyber jailbreaks on an internal benchmark and 94.8% across PolyGuard, RedCode and CyberBench 2024 with no system-level safeguards, versus 75.5% and 70.2% for Kimi K3 in the same setup. Chemical and biological refusals are above 98% on SecureBio's BioTIER, honesty on MASK is 98.1%, and the Agentic Misalignment misbehaviour rate is 0%. Meta is explicit that these results apply to the released checkpoint only and that fine-tuning can degrade them.

Comparison to prior work

The closest comparison is Meta's own history. Llama 3.1 405B, released in July 2024, was dense, so every token read all 405B parameters, which is why nobody ran it on a desktop. Llama 4, in April 2025, moved to MoE. Then Meta's rebuilt Superintelligence Labs shipped the original Muse Spark in April 2026 without weights, and on 12 October shipped Muse Wonder, also closed. Spark Open is the lab's return to publishing weights at frontier scale, with an architecture that looks closer to DeepSeek-style fine-grained experts (256 experts, top-8) than to Llama 4's coarser routing. Compared with Mistral Large 4, the other recent trillion-class open release, Meta's table has Spark Open ahead on the AA index (39 vs 38) and SciCode (58.5 vs 54.2) but behind on GDPVal-AA (1401 vs 1424) and DeepSWE (60 vs 62).

Cost

Hosted pricing at Together AI is $1.25 per million input tokens and $4.25 per million output, with $0.15 for cached input. That is slightly below GLM 5.3's $1.40/$4.40 and well above MiMo-V2.6-Pro at $0.435/$0.87. For the local route, the published figure is $10,799 per 256GB Mac Studio with 1TB storage, so four come to $43,196 before tax; the 512GB price is unannounced.

Why it matters

For everyday users the direct effect is nil; nobody is buying two Mac Studios to chat. The indirect effect is that the model is now on OpenRouter, Together, Groq, Fireworks and others, so apps built on it will appear quickly, and the "Built with Muse Spark Open" attribution clause means you will see the name.

For developers and builders this is the first time a US lab has published weights for a model with a 1M-token context, four input modalities and agent-tool scores in the same band as Kimi K3 and GLM 5.3. The practical change is that a team with sensitive code or footage can run a frontier-class agent with no data leaving the building, and Apple's post ends with exactly that recipe: an OpenAI-compatible endpoint on localhost that tools like OpenCode can point at. The published architecture, including the compressed-MoE block and the 16/46 attention split, is also a reference others can copy.

For companies the licence is the story. The $20M threshold targets cloud resellers, not users. A bank that fine-tunes the model for internal use owes Meta nothing beyond attribution. A startup that sells API access to it and crosses $20M in a year must negotiate. Trending Topics reads this as unlikely to hinder most developers while leaving Meta as gatekeeper for the hyperscalers. For Apple, the deal gives the Mac Studio a flagship workload: Apple's own post uses the model to showcase its new RDMA-over-Thunderbolt stack, and the company has not had a credible high-end AI inference story until now.

For the field the geopolitical framing is explicit. VentureBeat cites a Hugging Face assessment that Chinese models took 41% of downloads over the year to March 2026, overtaking US models, and OpenRouter data showing open-weight models at 60% of US-originated token volume, most of it from Chinese labs. Nathan Lambert of Interconnects wrote on X, "I waited so long for US open weight models to be relevant again," adding "Still more work to do!" That second sentence is the honest summary: Meta has re-entered the race, not won it.

Second-order effects: expect a wave of distillations, since the licence expressly permits distilling, and expect the two-Mac recipe to become a benchmark that Nvidia's DGX Spark and AMD's Strix Halo ecosystems respond to. One Hacker News commenter noted that GLM 5.3 Flash already runs at 50-60 tokens a second on two DGX Sparks.

What to be skeptical of

The "most capable US open-weight model" claim is narrow. It is true against Inkling and Nemotron 3 Ultra, and false against the open-weight field as a whole. By Meta's own table, Kimi K3, GLM 5.3 and Qwen3.8 beat it on coding and on the composite index. The model card does not include MiMo-V2.6-Pro, which Artificial Analysis scores at 46, highest of all. And the "US" lead is only as wide as it looks because Reflection AI's Beam, which Meta's table shows ahead on Terminal-Bench 2.1 (80.1 vs 76.4) and GPQA Diamond, has not yet published its weights.

Almost every benchmark number is vendor-reported. The one independent check so far is Artificial Analysis, whose own listing confirms the 39 Meta reported, at a cost of $0.84 per task. Everything else in the table is Meta's measurement, and several agentic coding results were run in each lab's own harness. The full safety and evaluation report is still "available soon".

Early hands-on reports are lukewarm. In the Hacker News thread, which had passed 690 points and 370 comments by the time of writing, user AustinDev reported benchmarking it locally against GLM 5.3 Flash and finding it "slower in TTFT, tok/s and concurrency" and lower-scoring. User a11r argued Meta "has a bit of catching up to do in the open-weight space". Others noted that on some coding benchmarks it lands between Meta's own closed Muse Spark 1.2 and 1.3. Another commenter observed that evaluations tend to run at native precision while people who care about speed run quantised builds, so the Mac results and the benchmark results describe different artefacts.

The Mac numbers are narrower than the headlines. Apple's 25 and 37 tokens a second were measured with an 8,192-token prompt generating 128 tokens. On a 65,536-token prompt the four-machine time-to-first-token was 40.6 seconds. The 1M context will not practically fit alongside the weights in 1TB. The two-machine configuration cannot be bought yet.

"Open" has strings. The weights sit behind a Hugging Face gate that asks for full legal name and date of birth. The licence requires attribution, bans certain uses and imposes the $20M model-as-a-service threshold. One commenter noted that Kimi K3's licence "contains an identical carve out"; another called it "a fair license" with "a bit of 'Open'-washing". It is not open source, and Meta does not claim it is.

Conflicts of interest. Baseten, whose technical note is the only source for the compressed-MoE and hybrid-attention details, sells hosted access to the model and served it from day zero. Apple's benchmark post is marketing for the M5 Ultra. Meta chose which models and benchmarks appear in its table.

What to watch next

• Independent agentic and coding evaluations. Artificial Analysis has already confirmed the 39 on its index. The open question is the agent-tool rows, where Meta's lead is widest and the only numbers so far are Meta's own.
• The 512GB M5 Ultra Mac Studio. It ships in late October with no announced price. The two-machine demo depends on it; until then the only working desktop recipe is four 256GB units.
• Meta's full evaluation report. The announcement promises one "soon". The Superintelligence Scaling Framework results and the full safety methodology are where the 0.5 StrongREJECT score and the 0% misalignment rate should be explained.
• Community quantisations and distillations. Unsloth and llama.cpp are supported engines at launch; watch for 3-bit and 2-bit GGUF builds that squeeze it onto a single 512GB machine, and for small distilled models, which the licence expressly permits.
• Microsoft's MAI-Base-2. Microsoft said the same day it had finished pre-training a 2.8T-total, 80B-active model, still in post-training, with weights to be licensed to "trusted partners" through Foundry rather than published. Whether Microsoft publishes weights would reshape the US open-weight picture again.
• How this fits Meta's two-track strategy. Four days ago Meta launched closed Muse Wonder at a higher independent score. Spark Open is the open track. If future open releases keep arriving one tier below the closed ones, the Llama-era promise of open parity is over, and this is the new deal.

How this was made: researched and written by an AI model (Claude) from the primary sources listed above, then checked claim-by-claim against those sources in a separate AI fact-check pass. Spotted an error? Email [email protected] and we correct it publicly. Our process.