Yesterday Meta released the weights of Muse Spark Open — 816 billion parameters — and the headline deployment story was not an API endpoint. It was a desk. Two Apple M5 Ultra Mac Studios, the MLX framework, and what Meta calls the most capable US open-weight model to date, running agentic coding workflows entirely on-device. Within 24 hours, Hugging Face had independent numbers: on a single M3 Ultra with 512 GB of unified memory, a 3-bit quant generates at about 21 tokens per second with roughly 225 tokens per second of prefill, occupying around 344 GB. That is not a benchmark. That is a receipt.
Most coverage filed the release under "Meta rejoins the open-weight race," which misses the actual event. For the first time, a US lab shipped near-frontier weights where the download is the deployment: day-zero support in llama.cpp, transformers, and MLX, pre-built quantizations, open serving kernels, a fine-tuning path, and a reference hardware configuration you can buy at retail. The interesting question stopped being whether open weights can compete. It's what happens to agent economics when the marginal token costs electricity instead of $10 per million.
What Actually Shipped
The specs first: Muse Spark Open is a natively multimodal mixture-of-experts model — text, images, audio, and video in; text out — with 816B total parameters, 42B active per token, and a 1M-token context window. Benchmarks we'll get to. What makes this release structurally different is everything that shipped around the weights:
- Day-zero inference support in llama.cpp, transformers, oMLX, and mlx-lm, with pre-built 3-bit and 4-bit "oQ" quants ready to run.
- Custom, auto-tunable Helion GPU kernels merged into SGLang, delivering 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs (3% end-to-end on AMD MI350). Meta open-sourced the serving stack, not just the model.
- A post-training path via TorchTitan, NVFP4 weights for production serving, and a single-command NIM container.
- Weights mirrored across Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry; API serving from day one at Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together.
Labs used to dump weights on Hugging Face and let the community spend six weeks making them runnable. Meta shipped a deployment kit. That's the behavior of a company that wants this model deployed, not merely downloaded and benchmarked.
And the proof-of-work demos were chosen deliberately: Meta says it ran full agentic workflows on-device — shipping real iOS changes to Ice Cubes, an open-source Mastodon client, from bug fixes through full build — and driving DaVinci Resolve end to end, reviewing footage frame by frame to find strong moments and crop them vertically. Vendor demos are vendor demos. But the shape of the demo is the tell: they showcased desktop workflows, not cloud dashboards.
42 Billion Is the Number That Matters
Buried under the 816B headline is the figure that determines everything about deployment: 42B active parameters. In a mixture-of-experts model, total parameters are what the model knows; active parameters are what it computes per token. Every token passes through roughly a 42B model's worth of compute while drawing on 816B of learned capacity. That single architectural fact is why any of the rest of this is possible.
It's why a desk can serve the model at all. Hugging Face's tests show an imatrix-calibrated 2-bit quant running at 18 tokens per second with 304 GB in use — and the 3-bit build at 21 tokens per second using 344 GB. The tiny speed difference between 2-bit and 3-bit tells you the bottleneck isn't compute, it's memory bandwidth. That's the physics of local inference, and it means the practical constraint on running frontier-class weights at home is unified memory capacity, not FLOPS.
The same weights on NVIDIA GB300 NVL72 racks — the hardware Meta trained the model on — deliver up to 50K tokens per second per GPU in aggregate and up to 600 tokens per second per user. Identical model, radically different economics, chosen by where you plug it in.
For agents specifically, two more details matter. The model ships with goal conditioning and context compaction baked in — machinery for holding direction across long-running sessions, which is an agent feature, not a chat feature. And the 1M-token context has a quiet financial implication: agents replay context. A coding agent re-sends its system prompt, repository map, and transcript on nearly every turn. Cloud vendors know it — GPT-6.1 Sol's cache reads now run 95% below fresh input pricing precisely because agent workloads demanded it. Locally, the KV cache sits in your RAM and nobody bills your RAM.
The Arithmetic of Owning vs Renting
Run the numbers on a concrete workload: one long-horizon agent session — say a code migration — that accumulates 1M tokens of context and generates 500K tokens of output.
- GPT-6 Astra ($10/$50 per million): $35.00 for the session.
- Muse Wonder, Meta's closed mid-tier ($4/$20): $14.00.
- GPT-6.1 Sol ($2/$10): $7.00, before any cache discounts.
- Local: 500K tokens at 21 tok/s takes 6.6 hours. Assuming ~200W under sustained inference load, that's about 1.3 kWh — roughly $0.22 of electricity at US residential rates.
Two orders of magnitude between the flagship API and the wall socket. Now the honest column of that ledger: wall clock and capital. Those 500K tokens take 6.6 hours locally versus about 14 minutes at the 600 tok/s per user that NVIDIA's serving stack delivers. And a 512 GB M3 Ultra is a serious machine whose price amortizes over years — the API bill compounds instead. One more honesty note: what runs on the desk is not the model, it's 3 bits of the model. Quantization loss is real and task-dependent.
So the verdict splits cleanly by workload. Interactive agents, where a human is waiting: rent, and rent the fast tier. Batch agents — overnight migrations, cron-triggered workflows, CI pipelines, eval sweeps, the 3 a.m. retry loops that nobody watches — are a different species. Nobody cares if the agent finishes at 4:11 a.m. instead of 9:40 p.m. For every batch workload you own, the calculation has collapsed to a capex decision, and the API price is no longer the ceiling it was eighteen months ago.
Read the Claim — All of It
"Most capable US open-weight model to date." Two qualifiers, both load-bearing.
First, US. On BenchLM's October leaderboard, Muse Spark Open scores 62.66. The best open-weight model on that same board is Xiaomi's MiMo-V2.6-Pro at 73.62 — an estimated ranking, but from the same methodology that scored Spark. By that measure, the open-weight crown still sits in China, and Meta's claim is a national title, not a world title.
Second, open-weight. Meta's actual best model — Muse Wonder Full, released three days earlier at 74.27 and holding BenchLM's October crown — is proprietary and available only through a Muse Code subscription. The gap between what Meta gives away and what Meta sells is 11.6 points, and that gap is a product decision. The download gets you last week's frontier. This week's costs $4/$20 per million tokens (Muse Wonder, the API tier between them) or a subscription.
The benchmark spread tells the same cautionary story. On DeepSWE v1.1, a long-horizon coding benchmark, Muse Spark Open scores 60.0% in the Muse Code harness per Meta's scorecard — solid, and 15 points behind GPT-6.1 Sol's 75.2%. On Terminal-Bench 2.1, Artificial Analysis measures it at 76.4%. On Terminal-Bench 4.0, from the same evaluator in the same week: 8.1%. Same model, adjacent benchmark versions, 68 points apart. A third-party estimate puts SWE-bench Verified at ~89.7% — an estimate, flagged as such, not a measurement. Its strongest verified agent numbers are MCP Atlas at 89.2% and Toolathlon-Verified at 72.8%.
The takeaway isn't "the numbers are fake." It's that any single score now tells you less than ever about a model you're about to wire into production. The only benchmark that decides anything is the one you run — on your workload, on your quant, in your harness.
Two Bets on Where Agents Live
Zoom out and this week becomes a clean natural experiment, with three companies answering "where does the agent run?" in three contradictory ways.
OpenAI answered at DevDay with Dots: always-on agents powered by GPT-6 Astra, each with its own cloud computer and persistent memory, pursuing user-defined goals between conversations — behind a new $500-per-month tier. Early reception has been rough; critics call the first rollout lazy and confusing. But the direction is unmistakable: agents as a hosted service, always-on in OpenAI's cloud, on OpenAI's meter.
Microsoft answered Thursday in Copilot Studio: agent accounts and business process memory, plus triggers in the GitHub Copilot harness that let an agent wake on an email or a cadence with no prompt at all. These are managed agents that "belong to the organization" and "persist beyond any one individual's role or tenure." The enterprise answer: agents as governed staff, running in the tenant.
Meta answered with 816 billion parameters and a suggested pair of Mac Studios. The appliance answer: own the runtime, own the weights, own the electricity bill.
Notice that "always-on" is precisely the workload where owned hardware wins. An idle agent on someone else's GPU is a metered liability; an idle agent on your desk is a space heater you already paid for. OpenAI is betting you'd rather rent time than own hardware. Meta is betting the opposite. Microsoft is betting large organizations want neither — they want audit trails. All three can be right for different customers, but every agent architecture decision you'll make next year — where state lives, how context gets cached, what triggers a run — now has a property line drawn under it.
What to Do This Week
Price your batch workloads per completed task, not per token. Take your nightly and weekly agent jobs and compute cost-per-task at API prices versus owned hardware, using your actual wall-clock tolerance. For jobs with no SLA, the crossover has likely already passed you. And remember the middle path: Spark Open is served by Groq, Fireworks, Together, OpenRouter and others — open weights on rented silicon, swappable by config, with no vendor owning your stack.
Rerun your own evals before quoting any chart. When the same model scores 76.4% and 8.1% on adjacent Terminal-Bench versions the same week, vendor scorecards are directional at best. A 200-task sample of your real work, in your harness, on the quant you'll actually run, beats every leaderboard on the internet.
Test the quant, not the model. If you go local, the artifact you ship against is the 3-bit or 2-bit build on your specific hardware. Quantization damage is uneven across tasks — it may be invisible on summarization and fatal on precise tool-calling arithmetic. Measure where it hurts before you commit.
Keep model identity in configuration anyway. Downloadable weights still get patched, re-quantized, and superseded — Muse Glimmer in August, Spark Open now, something else next quarter. The core discipline of harness engineering — model, effort level, and serving setup as versioned config rather than hardcoded assumptions — applies to a model running on your desk exactly as it does to one rented by the million tokens.
Whose Electricity
The frontier you can download now sits roughly one release behind the frontier you rent — 11.6 points by BenchLM's board — and it runs, slowly and cheaply, on hardware you can order today. That gap will not hold constant. Meta has every incentive to keep shrinking it, because their business is engagement, not tokens, and every agent running on a desk somewhere is an agent not paying OpenAI's rates.
Each quarter the gap narrows, the question "which API should we use?" gets less interesting than the question it was always hiding: whose machine does your agent live on? This week, for the first time, "yours" became a defensible answer.