Oct 16, 2026

Meta Muse Spark Open 816B for Local AI: The 256GB Floor, Two Mac Studios You Can't Order Yet, and What Actually Runs It (2026)

By RunAIHome Team · 13 min read

metamuse-spark-openlocal-llmmac-studiohardware-guide

TL;DR: Meta shipped Muse Spark Open on October 15 — an 816B-parameter, 42B-active multimodal MoE with a 1M-token context, billed as the most capable US open-weight model to date. It has day-one llama.cpp support and official GGUFs, but the smallest usable quant needs roughly 256–300GB of memory, and Apple’s flagship demo runs on two 512GB M5 Ultra Mac Studios that aren’t orderable yet.

	2× Mac Studio M5 Ultra 512GB	1× 512GB machine, 2-bit	OpenRouter API
Best for	The full 4-bit model, fastest local path	Proving it runs on one box	Actually using it today
Price / Cost	Not orderable yet (512GB tier ships late Oct, price TBA)	Used M3 Ultra 512GB: ~$18K+ asks (Sep 2026)	$1.25/M in, $4.25/M out
Speed	24.99 tok/s (Apple, tensor parallel)	~18 tok/s (Hugging Face, M3 Ultra)	Provider-dependent
The catch	You cannot buy the hardware yet	Discontinued machine at a collector premium	It’s not local

Honest take: This release matters for what it signals — Meta putting a frontier-scale MoE into llama.cpp on day one — not for what you should buy. No consumer GPU path exists, the Mac path starts at a machine Apple hasn’t priced, and the benchmark table says Kimi K3 and GLM-5.3 are still ahead anyway. Use the API, keep your 24GB card on Nemotron 3.5 Bolt or Qwen3.6, and revisit if the 512GB M5 Ultra lands under $12K.

Meta’s open-weight story in 2026 has two chapters. August’s Muse Glimmer 30B was the consumer chapter: Apache 2.0, 17.3GB Q4 on a used RTX 3090. Muse Spark Open is the flag-planting chapter: 816 billion total parameters, 42 billion active per token, text/image/video/audio input, a 1,048,576-token context window, and a launch pitch aimed directly at the Chinese open-weight leaders — “the most capable US open-weight model to date,” per Meta’s announcement.

The qualifier is doing real work in that sentence, and we’ll get to it. First, the question this site exists to answer: what does it take to run this thing, and should you?

What shipped on October 15

Per the Hugging Face launch post and the model card:

Architecture: sparse mixture-of-experts, 816B total / 42B active per token. For decode speed, the 42B active figure is what matters — each generated token reads roughly 5% of the weights.
Modalities: text, image, video, and audio in; text out.
Context: 1,048,576 tokens (the OpenRouter listing confirms the full 1M window is served).
Formats: BF16 and NVFP4 on Hugging Face (gated — you accept the license before downloading), plus official GGUFs from ggml-org in Q2_K, Q4_K_S, and Q8_0, plus Unsloth Dynamic quants. llama.cpp support was merged for launch day and ships in the prebuilt binaries.

That last line is the quiet headline. Mistral Large 4 was announced October 6 with weights promised “end of October.” Reflection’s Beam 501B is still in final red-teaming. Meta shipped weights, GGUFs, and merged llama.cpp support in one day. Whatever you think of the model, that’s how an open-weight launch should work.

The license is not Glimmer’s license

Glimmer was plain Apache 2.0 — Meta’s first. Spark Open ships under the Muse Model License, a custom license with three terms that matter for home-lab and indie use, per FourWeekMBA’s breakdown and SiliconANGLE:

The $20M clause targets API resellers, not you. If you sell hosted access to the model as a model-serving business and you (plus affiliates) pass $20 million in gross revenue over any 12-month window, you must notify Meta and sign a separate agreement. Internal use, research, and using the model inside your own product are explicitly outside the clause.
Attribution is required. Products built on it must display “Built with Muse Spark Open” somewhere permitted — docs, website, or settings count. Fine-tunes and distills inherit the duty.
Outputs are yours. Meta claims no ownership of generations. Military and critical-infrastructure uses are prohibited.

For a homelab, none of this bites. It is not Apache 2.0, and anyone writing “fully open source” is wrong — but the practical answer for a self-hoster is: you can run it, fine-tune it, and build on it without paying anyone.

The memory wall, tier by tier

Here are the real numbers, from the model card, ggml-org’s repo, and Unsloth’s run guide:

Format	Size on disk	Memory to actually run it	Who can load this
BF16	~1.6TB	~2TB per the model card	Nobody at home
Q8_0 (ggml-org)	~900GB	~1TB class	Nobody at home
NVFP4	—	~700GB VRAM per the model card	Server racks
MLX 4-bit	~493GB	>512GB (doesn’t fit one 512GB Mac)	2× Mac Studio 512GB
Q2 class (Unsloth UD-IQ2_XXS / ggml-org Q2_K)	~300GB class	~256GB RAM per Unsloth; Meta says ~300GB combined RAM+VRAM	One 512GB machine, comfortably; 256GB is the knife’s edge

Measure that bottom row against what exists. A 256GB M5 Ultra ($9,499) is exactly at Unsloth’s floor — workable for the weights, with the KV cache and OS fighting for scraps; Meta’s own ~300GB combined figure says treat 256GB as experimental, not comfortable. Framework’s 192GB Desktop tops out at a 160GB usable pool — out. A quad-3090 build’s 96GB — not remotely close. Even the 2-bit tier of this model starts above every configuration this site has ever recommended to its readers.

And 2-bit is the aggressive end of quantization. Our Kimi K2.7 guide covers the same trade-off at the same scale: you accept measurable quality loss to make a trillion-class model fit at all. The benchmarks below were not run at 2-bit.

The Apple demo: real numbers, imaginary hardware

The launch’s most interesting artifact is Apple’s own benchmark set. Meta and Apple coordinated this release — MLX gained RDMA over Thunderbolt 5, letting multiple Macs exchange tensors machine-to-machine, and Apple published throughput for the 4-bit model (~493GB) on 512GB M5 Ultra Mac Studios, per MacRumors and iClarified:

Setup	Parallelism	Decode speed
2× M5 Ultra 512GB	Pipeline	15.74 tok/s
2× M5 Ultra 512GB	Tensor	24.99 tok/s
4× M5 Ultra 512GB	Tensor	~37 tok/s

Sanity-check against the bandwidth ceiling, because vendor numbers that fail this test are fabrications: at 4-bit, each token reads about (42/816) × 493GB ≈ 25GB of weights. One M5 Ultra moves 1.2 TB/s, so a single machine’s hard ceiling is ~47 tok/s. Apple’s 25 tok/s across two machines — paying Thunderbolt interconnect tax on every layer — is 53% of a single-box ceiling. Plausible, even good. The same math on Hugging Face’s single-machine run (2-bit, ~304GB resident, M3 Ultra at 819 GB/s: ceiling ~52 tok/s) makes their measured ~18 tok/s equally believable.

Now the part the launch coverage kept skipping: you cannot buy a 512GB M5 Ultra today. Apple shipped every other M5 Ultra configuration in September; the 512GB tier is “late October,” price unannounced, with TechRadar speculating a starting price well above $10,000. The only 512GB Apple Silicon you can get this week is a discontinued M3 Ultra at ~$18,000+ used asks (September eBay, per our M5 Ultra vs M3 Ultra breakdown) — a collector premium for the slower chip. “Runs on two Mac Studios” is true, and for at least a few more weeks it describes a computer that exists only in Apple’s lab and Meta’s launch video.

When the 512GB tier does open: two of them, at a speculative $12K each, is a ~$24K rig producing 25 tok/s. DeepSeek V4 Flash via DS4 gets comparable agent-class capability onto one 128GB machine. Scale has a price curve, and it is not linear.

What about the NVIDIA path?

There isn’t a consumer one. The 2-bit floor (~300GB) is 13 RTX 3090s of VRAM; nobody builds that on PCIe risers for inference sanity. Four RTX PRO 6000 Blackwell 96GB cards = 384GB and clears the 2-bit and even approaches MLX-4-bit territory — at $13,998–$19,999 per card (street, October 2026), a $56K–$80K GPU bill. The CPU route is equally theoretical this year: llama.cpp will happily run the Q2_K from system RAM with --cpu-moe, but 512GB of DDR5 RDIMM costs ~$7,200+ in the current DRAM crisis (64GB RDIMMs run $900+ each), and our Kimi K2.7 coverage pegged a 384GB EPYC build at 8–11 tok/s for a 32B-active model — Spark Open reads ~50% more per token (42B active at a slightly larger quant), so expect single digits. Estimated, not measured; the direction is not in doubt.

Trying it anyway: the gated-repo trap

The GGUFs are the one piece a mortal can actually touch — ggml-org’s quants pull straight into llama.cpp:

# Will download ~300GB. Check your disk first.
llama-server -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q2_K --cpu-moe -c 8192

The error you’ll likely hit first has nothing to do with hardware:

error: failed to get manifest ... 403 Forbidden

Muse Spark Open is a gated release — the base repo requires logging in to Hugging Face and accepting the Muse Model License on the model page before any download token works. The fix: open the model page in a browser, accept the terms, then authenticate your CLI (hf auth login, or export HF_TOKEN) before retrying. The same 403 bites huggingface-cli download and Ollama imports; it’s the license acceptance, not your network. (Mirrors of the ggml-org quants may skip the gate — per our usual rule, treat third-party re-uploads of day-one weights as unverified until checksums match.)

If the API is how you end up using it, it plugs into coding tools the same way any OpenRouter model does — the BYOK setup in aicoderscope.com’s local-backend guide applies unchanged, just with a hosted endpoint. And if this release is your first push into self-hosting anything, start smaller: aifoss.dev’s Ollama review covers the stack that runs the models your hardware can actually hold.

Set context expectations honestly, too: the 1M window is an API-tier feature. KV cache at frontier context costs hundreds of gigabytes on top of weights; on a 256–512GB local box you’ll run 8K–32K context, same as every other trillion-class local experiment we’ve covered.

The benchmark asterisk

Meta’s launch table leads with agentic results: 89.2 on MCP-Atlas, the top score in its own comparison, per MarkTechPost. The rows Meta didn’t lead with, from the same coverage plus AI-TLDR’s roundup:

Benchmark	Muse Spark Open	Best open rival
MCP-Atlas (agentic tool use)	89.2	— (Meta’s table, vendor-reported)
Terminal-Bench 2.1	76.4	Kimi K3: 85.0
Artificial Analysis Intelligence Index	39	GLM-5.3: 45, Kimi K3: 44

“Most capable US open-weight model” is accurate and carefully worded. On the independent composite index, two Chinese open-weight models still score higher; on terminal agent work, Kimi K3 leads by 8+ points. Spark Open’s genuine edge is the multimodal input stack plus the 1M context plus MCP-native tool use in one open checkpoint — a combination nothing else open currently offers. Independent verification hasn’t landed yet (AlphaSignal flags the same), and every number above is vendor-reported except the AA index.

A naming note before the FAQ, because October 2026 is cursed: Meta’s Muse Spark family has no relationship to NVIDIA’s DGX Spark desktop or the RTX Spark laptop chip. Three products, one word, zero shared silicon.

What to actually buy

Prices as of October 2026, verified against the comparison above:

Your situation	The move	Price	Where
You want to use Spark Open, starting now	OpenRouter API (Together, Fireworks, Nebius, Baseten)	$1.25/M in, $4.25/M out, $0.15/M cached	OpenRouter
You want frontier-class local AI and own nothing yet	Used RTX 3090 24GB and run 70B-class models that fit	$1,399–$1,450	Check price
You already own a 24GB card	Run Nemotron 3.5 Bolt 74B — shipped the same day, fits your card	$0	—
You’re determined to run Spark Open at home	Wait for 512GB M5 Ultra orders to open (late Oct, price TBA) — don’t pay $18K+ for a used M3 Ultra	TBA	—
You want to test the 816B before committing anything	Rent a multi-GPU instance by the hour	pay per hour	Vast.ai

FAQ

Can my 24GB (or 32GB, or 128GB) machine run Muse Spark Open? No. The smallest published quant needs ~256GB of memory at the absolute minimum, ~300GB per Meta’s combined figure. That excludes every consumer GPU, every Strix Halo machine, and every Mac below 512GB (the 256GB M5 Ultra is borderline-experimental). SSD-streaming engines don’t rescue it either: Strata is hard-wired to one 125B model, and generic SSD offload at 42B active parameters per token lands in the low single digits of tok/s.

Is the license safe for a homelab or a small SaaS? Yes. The $20M revenue clause applies to businesses selling hosted access to the model itself, and the threshold is gross revenue across your whole business — but internal use, research, and powering features of your own product are exempt regardless of revenue. You do owe a visible “Built with Muse Spark Open” if you ship something on it.

Is it the best open-weight model? Vendor-reported agentic scores say it leads on MCP-style tool use; the independent Artificial Analysis index (39) places it below GLM-5.3 (45) and Kimi K3 (44). It is the best US open-weight release and the only open frontier model with video+audio input and 1M context. “Best overall” remains a Chinese two-horse race until independent evals land.

Two Mac Studios at ~25 tok/s — is that actually good? Against the physics, yes: ~53% of a single machine’s bandwidth ceiling while paying Thunderbolt 5 tax is efficient distributed inference — far better than the EXO-era pooling that added capacity but rarely speed. Against your wallet, no: ~$24K (estimated) for 25 tok/s when the hosted API serves the same weights at $1.25/M input.

Will there be a consumer-sized Spark? Nothing announced. Meta’s consumer open-weight slot is the Glimmer line (30B dense, Apache 2.0). If a Spark distill follows the pattern of this fall’s other releases, that’s the version this site would tell you to buy hardware for — the 816B is infrastructure, not a homelab target.

Sources

Last updated October 16, 2026. Prices and specs change; verify current rates before purchasing.

Was this article helpful?

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.