Apple runs an 816B open model on four Macs; the case is control, not cost

*Four desktops reach 36 tokens a second on Meta's Muse Spark Open, which buys privacy and barely dents the API bill.*

Opening read

Apple's researchers have run Meta's 816-billion-parameter Muse Spark Open across four Mac Studios at better than 36 tokens a second, and my read is that this is the strongest evidence yet that frontier-scale open models can live on hardware you own, while leaving the economics unflattering. At that speed one cluster would displace roughly $13 a day in output-token fees against the hosted price of Muse Spark, the closed sibling, so the stronger reason to buy is control over data and vendors. The roundup covers Reflection's Beam landing in the open tooling and Google's self-correcting lab robots.

Local inference: four Macs run a frontier-scale open model, with a catch

This brief has spent the past week tracking whether real AI work can run on computers a company owns instead of rents, and the honest yes answers so far have involved small models. Apple's research group has now moved that ceiling a long way: it ran Meta's Muse Spark Open, an 816-billion-parameter model (the size of a lab's top tier), across four Mac Studios with M5 Ultra chips and got more than 36 tokens per second, the word fragments models emit Apple Machine Learning Research (https://machinelearning.apple.com/research/mlx-distributed-inference). The consensus read writes itself: a cloud GPU rack is no longer the only home for a model this large, so buyers should start pricing desktops against rentals.

That read holds for two reasons: the model's shape cuts the work per word, and the new chips do that smaller job faster. Muse Spark Open is a mixture-of-experts model, meaning only about 42 billion of its 816 billion parameters do work on any given token, so the machines must hold everything in memory but compute only a slice per word, and the M5 Ultra Mac Studio can be configured with up to 512 GB of memory, the kind of room a four-way split of a model this size needs Hugging Face Blog (https://huggingface.co/blog/meta-muse-spark-open) Nate Jones (https://www.youtube.com/watch?v=1lO8aNSLPJc). The Macs supply the memory to hold the whole model, and the model's design keeps the work per word small. Apple's researchers tied the four boxes together with MLX (Apple's software for running AI models on its own chips) and RDMA (a way for one machine to read another's memory directly), a capability the exo project says Apple promotes on its product pages as low-latency RDMA over Thunderbolt 5 Apple Machine Learning Research (https://machinelearning.apple.com/research/mlx-distributed-inference) Latent Space (https://www.latent.space/p/ainews-hot-chips-openais-jalapeno). The hardware then turns that fit into usable speed: Apple attributes an up to 4.6 times shorter wait before the first word to new accelerator units on the graphics chip, and 1.6 times faster generation to memory bandwidth, both measured against the older M3 Ultra Apple Machine Learning Research (https://machinelearning.apple.com/research/mlx-distributed-inference). Those are generational jumps, and they are the second half of the answer.

Here is where I part from the buy-now read. Thirty-six tokens per second, run flat out for 24 hours, is about 3.1 million tokens, the most output a single cluster could produce at Apple's reported rate. Priced against hosted Muse Spark, the closed sibling, at $4.25 per million output tokens, that would displace roughly $13 a day in output fees, and at Meta's $0.20 contributor tier the figure drops well under a dollar Big Technology (https://www.bigtechnology.com/p/what-i-got-wrong-about-metas-muse) Simon Willison's Weblog (https://simonwillison.net/2026/Aug/5/muse-code-and-muse-spark-12/#atom-everything). Treat that as a hypothetical rather than a verdict. The figures available here do not say whether 36 tokens per second is one conversation or several served at once; batching would raise the ceiling, and I do not know by how much. The supplied pricing evidence gives no hosted price for Muse Spark Open itself, even though the same open weights can now be rented through Amazon SageMaker JumpStart, and nothing here prices the four-Mac cluster, so any saving has to be set against your own hardware quote AWS Machine Learning Blog (https://aws.amazon.com/blogs/machine-learning/metas-muse-spark-open-now-available-in-amazon-sagemaker-jumpstart/). That rental option actually sharpens the real case, which is control: a hosted deployment, SageMaker included, still sends your documents off site, while a cluster you own means no rate limits, no surprise repricing by a vendor, and data that never leaves the building, which is why confidentiality-bound professional services firms are already buying Mac Minis to run local models Nate Jones (https://www.youtube.com/watch?v=RaAFquzj5B8). For an IT leader deciding whether to fund a Mac cluster for a pilot, that turns the proposal from a savings line into a risk-reduction line, and it should be argued that way in front of whoever signs.

A check before you write it up: pull the output-token count for one workload from your API dashboard for the last seven days and divide by seven. Take the smaller of that daily figure and 3.1 million, because 3.1 million is the ceiling one cluster implies at Apple's rate run nonstop, with the serving conditions behind it unknown. Multiply the capped figure by your current vendor's output rate per million tokens, using $4.25 only if the comparison is hosted Muse Spark, and you have the most a cluster could save per day in output fees. Set the saving against your hardware quote; if it looks small next to that quote, argue the proposal on control instead.

Open models: Beam lands in the tooling, with a disputed score gap

Reflection AI launched Beam on October 5 as a US-trained open-weight (downloadable model weights) answer to Chinese models, and this week the plumbing caught up: Hugging Face published its welcome post and shipped native Beam support in Transformers 5.20.1, a patch that also repairs a 5.20.0 break in loading models on older versions of PyTorch, the underlying math library TechCrunch AI (https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/) Hugging Face Blog (https://huggingface.co/blog/reflection-beam) Transformers Releases (https://github.com/huggingface/transformers/releases/tag/v5.20.1). Pin versions before you pull. Beam fires 23 billion of its 501 billion parameters per token, and its 1 million token context (how much text the model can read at once) comes from a mix of layers, 43 that look only at nearby text and nine that scan the whole document Hugging Face Blog (https://huggingface.co/blog/reflection-beam). Whether that mix keeps memory in check at full length is a thing to test, not assume. The score picture is disputed: Reflection says Beam scores as well as the best Chinese open models on hard reasoning tests while costing less to run, while Import AI's read is that it trails the Chinese model GLM-5.3 on benchmarks and leans on a token-efficiency claim, meaning it reaches the same answer while generating fewer word fragments TechCrunch AI (https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/) Import AI (https://importai.substack.com/p/import-ai-476-creativity-and-rsi). For an engineer picking an open model for long-document agents, that efficiency claim is the thing to measure on your own corpus before either side's score story decides for you.

Lab robotics: Google's agents now fix their own experiments

Google Research has built VERID, a digital twin (a live software copy of the physical lab) on the open-source PyLabRobot, so an AI agent can dry-run and safety-check a procedure before a robot touches a pipette, and in aptamer experiments (short DNA strands selected to bind a target) confirmed hit rates, the share of designed candidates that actually bound the target, reached 33.5% and 42.1% after the agents spotted and corrected their own wet-lab failures Google Research Blog (https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs/). An earlier AI co-scientist handed experts drug targets that held up at the bench; the new turn is the agent running the bench itself AI Engineer (https://www.youtube.com/watch?v=wNH3q9pqn0U). This is the most concrete lab-robotics result in the day's evidence because it reports wet-lab corrections and measured hit rates rather than a demo. The stated goal, protein binders that grip cancer targets tightly enough to matter for drugs, is a plan rather than a result. For a life-sciences leader weighing lab-automation spend, the number to demand is the hit rate before the agents' corrections; Google's wording implies one exists, but this brief does not have it.

Augur calls

• Call: Google Research will publicly report a protein binder with single-digit nanomolar affinity for a named cancer target that was produced by its closed-loop AI agent and wet-lab robotics system by 2027-06-30.
• Criteria: A Google Research or Google DeepMind blog post, paper, or preprint that names the target and reports a measured affinity under 10 nanomolar for a binder designed and tested through the closed-loop system counts as a hit; no such public report by the due date is a miss.
• Due: 2027-06-30
• Odds: 45%
• Call: Reflection AI's Beam will be listed in Ollama's public model library by 2026-12-31.
• Criteria: A Beam entry on the public ollama.com library pages, or an Ollama release note announcing Beam support, counts as a hit; absence of both by the due date is a miss.
• Due: 2026-12-31
• Odds: 55%

What to watch

• Independent long-context results for Reflection's Beam would test the token-efficiency claim against the disputed benchmark picture. Import AI (https://importai.substack.com/p/import-ai-476-creativity-and-rsi)
• Google Research has promised closed-loop discovery of tight-binding proteins; a published binder with a named target is the proof point. Google Research Blog (https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs/)
• The supplied pricing evidence gives no hosted price for Muse Spark Open; a listing at or below Muse Spark's $4.25 output rate, on SageMaker or elsewhere, would shrink the local cost case further. Big Technology (https://www.bigtechnology.com/p/what-i-got-wrong-about-metas-muse) AWS Machine Learning Blog (https://aws.amazon.com/blogs/machine-learning/metas-muse-spark-open-now-available-in-amazon-sagemaker-jumpstart/)
• Black Forest Labs has now shipped FLUX 3 Image Klein, the 5.6B open-weight model, after promising an open-weight variant at the October 1 launch; independent tests of its 4K generation and editing claims are the next check. Latent Space (https://www.latent.space/p/ainews-pi-10-pi-durable-and-aie-nyc) Black Forest Labs (https://bfl.ai/blog/flux-3-image-klein)
• Hugging Face's follow-on Transformers patches should show whether the older-PyTorch loading break in 5.20.0 was fully closed by 5.20.1. Transformers Releases (https://github.com/huggingface/transformers/releases/tag/v5.20.1)

Verification notes

Apple's 36 tokens per second is a vendor-reported figure, and the figures available here do not say whether it covers one conversation or many served at once, which changes the cost arithmetic above Apple Machine Learning Research (https://machinelearning.apple.com/research/mlx-distributed-inference). The 512 GB memory figure is the top M5 Ultra configuration, not a stated spec of Apple's test machines Nate Jones (https://www.youtube.com/watch?v=1lO8aNSLPJc). The $4.25 and $0.20 output prices are for hosted Muse Spark tiers, not the Open variant, and the daily-saving figures are hypotheticals built on them with no cluster price to set against; input fees ($1.25 per million on hosted Muse Spark) are left out because the evidence gives no figure for how fast the cluster processes input Big Technology (https://www.bigtechnology.com/p/what-i-got-wrong-about-metas-muse) Simon Willison's Weblog (https://simonwillison.net/2026/Aug/5/muse-code-and-muse-spark-12/#atom-everything). Reflection's parity claim for Beam is the vendor's own, and Import AI's gap read is one outside assessment. Google's hit rates come from its own writeup, the hit rate before the agents' corrections is not in the evidence here, and the binder goal is a forecast Google Research Blog (https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs/).

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-10-17?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-10-17)