Meta opens Muse Spark Open, and Apple runs it on Mac Studios.
Meta’s 816B Muse Spark Open ships open weights, and Apple shows it running across Mac Studios
The plumbing was ready on day one. Hugging Face Transformers and TRL shipped support for inference and fine-tuning, mlx-lm and llama.cpp added it within hours, SGLang and vLLM have reference integrations, and hosted APIs went live at Together, Fireworks, Groq, Baseten, Modal, Nebius, and OpenRouter. The weights sit on Hugging Face behind a manual-approval gate, in BF16 (about 1.6 TB) and NVFP4. Ollama lists it as a cloud model only.
The Mac part is the news for this desk. Apple says the 4-bit NVFP4 build needs about 493 GB for weights alone, so it will not run on a single Mac Studio once you add working memory. Instead, MLX splits it across machines over macOS’s RDMA-over-Thunderbolt 5 driver. Apple reports about 16 tokens a second on two M5 Ultras in pipeline mode, 25 with tensor parallelism, and 37 on four, at an 8K context. Hugging Face’s own tests found smaller 3-bit and 2-bit quants do fit one 512 GB M3 Ultra at roughly 18–21 tokens a second. All speeds are vendor- or partner-reported.
Two cautions. The benchmarks are Meta-reported, and Meta’s own table shows it trailing several of the open models it compares against on Terminal-Bench 2.1 and GDPVal-AA, so “most capable US open-weight model” carries a qualifier. The Muse Model License lets you use, modify, distill, and sell, but you must show “Built with Muse Spark Open” in your product, follow Meta’s usage policy, and sign a separate deal with Meta to offer it as a model service once revenue passes $20 million.
What to do: if your Mac Studio has 256 GB, this one does not fit. Hugging Face says even the 3-bit build needs more than that. Try it through a hosted API or Ollama’s cloud first, and only plan a second machine if it beats your current local model on your own agent evals.
Sources: Meta AI · Apple Machine Learning Research · Hugging Face blog · Model card
Cloudflare says a pilot sharing its network signals with OpenAI’s search crawler, run on about 1% of its network, helped OpenAI find and index hundreds of millions of previously unknown pages (Cloudflare-reported). Cloudflare
NVIDIA expanded automated-driving partnerships with Volvo Cars and Mercedes-Benz in back-to-back announcements. NVIDIA Newsroom
Model & runtime releases
| Item | Date | Badge |
|---|---|---|
| Muse Spark Open 816B-A42B (BF16 + NVFP4; gated) | Oct 15 | Open |
| MLX distributed inference over Thunderbolt RDMA (Apple recipe) | Oct 15 | Open |
| mlx-lm 0.32.1 / 0.32.2 (Muse fixes) | Oct 15 | Open |
| llama.cpp v0.7.0 + Muse Spark Open support | Oct 15 | Open |
| Ollama v0.40.5 (Nemotron 3.5 Super / Bolt) | Oct 15 | Open |
| Nemotron 3.5 Super VL 120B-A12B (OpenMDW) | Oct 15 | Open |
| FLUX.3 Image Klein (Kommunity license) | Oct 15 | Open |
| Transformers v5.20.0 · TRL v1.16.0 | Oct 15 | Open |
| IBM Granite Switch 4.1 (chat + decisions + speech, one endpoint) | Oct 15 | Open |
| Claude Opus 5.5 fast mode GA | Oct 15 | Cloud |
| GPT-Live native Decisions integration | Oct 15 | Voice |
| Liquid AI d1-ultra decision model (API) | Oct 15 | Cloud |
| ExecuTorch Cortex-M backend (MCUs) | Oct 15 | On-device |
| RTX Spark PCs on sale (up to 128 GB unified) | Oct 16 | Active |
Security / policy / access
| Item | Date | Badge |
|---|---|---|
| Google OAuth DPoP for refresh tokens | Oct 15 | Active |
| Muse Model License: attribution + $20M model-service clause | Oct 15 | Watch |
| FLUX Kommunity License: commercial output under $5M revenue | Oct 15 | Watch |
| Anthropic embeds METR evaluators (publication rights) | Oct 15 | Active |
| Claude Compliance API: user instructions endpoint (beta) | Oct 15 | Cloud |
| Cloudflare Gateway: Workers in the policy path + custom DLP | Oct 15 | Cloud |
| vLLM Nemotron-VL image-bomb advisory — be on 0.31.0 | Oct 9 | Active |
| No new llama.cpp / Ollama / MLX / SGLang advisories | Oct 15–16 | Quiet |
Critical
Shipping anything on Muse Spark Open? The license requires a visible “Built with Muse Spark Open” credit in your product or docs, and that applies to fine-tunes and distilled models too. Serving Nemotron-VL models on vLLM? Be on 0.31.0.
Watch
Meta’s full Muse Spark Open evaluation report is “coming soon.” NVIDIA’s Super VL card points to a vLLM 0.32.0 image that has no GitHub release yet. Nemotron 3.5 Bolt E4B (4 GB) is due Oct 23.
Opportunity
Apple’s recipe uses 512 GB M5 Ultras. Hugging Face’s 3-bit Muse quant is about 344 GB, so two 256 GB Studios are the obvious config nobody has published numbers for yet. Apple’s new tool-discovery study is also worth a read if your agent picks from hundreds of tools.
Benchmark your Mac Studio against Apple’s table
pip install -U mlx mlx-lm, then run mlx_lm.benchmark --model Qwen/Qwen3.8-27B -p 8192 -g 128 -b 1 and compare your time to first token and tokens per second with Apple’s M3 Ultra and M5 Ultra columns. Now you know what an upgrade would buy you.
Split one model across two Macs
With two Thunderbolt 5 Macs on macOS 26.2 or later, follow Apple’s hostfile setup and run mlx.launch --hostfile hostfile.json --backend jaccl -- mlx_lm.benchmark --model <a model that fits> -p 8192 -g 128 -b 1 --pipeline. Learn the recipe before you need it.
Kick Muse Spark Open’s tires without buying hardware
ollama run muse-spark-open:cloud (needs an Ollama account) and point it at one of your own agent tasks. If it doesn’t beat your current local model, you’ve saved yourself a second Mac Studio.
Thursday through Friday morning (Oct 15–16 ET). ASTGL / Fraser takes, sourced.
The frontier fits in a stack of Mac Studios, not in one
Meta’s headline says desktop. Apple’s fine print says two machines at minimum, at 512 GB each, linked by Thunderbolt 5. Both are true, and that gap is the story. The lasting win here is not this one model. It is that Apple has now documented, with commands and numbers, how to treat a few Macs as one inference box: pipeline mode for fit, tensor mode for speed. My take: if you own one 256 GB Studio, this is not your model, and that is fine. Learn the multi-Mac recipe on a model that already fits, because the next big open release will assume it.
Sources: Apple ML Research · Meta AI
“Most capable US open-weight model” is a qualified claim
Read Meta’s own benchmark table and you find the word “US” doing work. Muse Spark Open leads on some agentic rows and trails several of the open models it lists on Terminal-Bench 2.1 and GDPVal-AA. All of it is Meta-reported, and the full evaluation report is still “coming soon.” The license is friendlier than past Meta terms, but the attribution line and the $20 million model-service clause are real obligations. My take: treat this as a strong option for coding agents you host yourself, not a default. Run your own evals and have someone read the license before it touches a product.
Sources: Model card + Muse Model License
Fast mode is just a tier now
Anthropic took fast mode for Claude Opus 5.5 out of research preview. No beta header, a default rate limit in the Start, Build, and Scale tiers, and the same option on Claude in Microsoft Foundry. OpenAI did the same thing for GPT-6.1 Sol with Ultrafast last week. My take: latency is now a line item you buy, like context or caching. Put it only on the paths a human is actually waiting on, check the pricing page before you flip it on, and keep batch and background work on the standard tier.
Sources: Claude Platform release notes (Oct 15) · OpenAI API changelog
Review is where agent speed goes to wait
Claude Code Review is out of preview for Team and Enterprise. Anthropic says most reviews now take 6 to 10 minutes instead of about 20, and admins can pick Quick, Standard, or Deep per repository. Those speeds are Anthropic-reported. The pitch is right: coding agents made writing code cheap, and human review is the queue. My take: turn on Deep only for the repos that can hurt you, and track how many findings your engineers actually accept. A reviewer people learn to ignore is worse than none.
Sources: Claude
NVIDIA wants the desk, too
RTX Spark laptops went on sale this morning with up to 128 GB of unified memory, and NVIDIA shipped open Nemotron 3.5 Super VL (120B, 12B active) plus Bolt, a 74B model squeezed into 24 GB. Its “5x faster than the next best” line is NVIDIA-reported. My take: for local agents that fit in 128 GB, the Mac Studio now has real Windows competition. For anything that needs 256 GB or more, Apple still owns the desk. One housekeeping note if you serve Nemotron-VL models on vLLM: an October advisory says older versions turned off Pillow’s image-bomb guard. Be on 0.31.0.
Sources: NVIDIA Newsroom (Oct 16) · NVIDIA blog · vLLM advisory
Anthropic gives METR a badge
Anthropic says six METR evaluators now have company hardware, office space, and internal access close to its own risk teams, and Anthropic cannot veto what METR publishes, only redact narrow categories. My take: this is the right shape for outside oversight, and it gives buyers a new question for every model vendor. Who outside your company can see how you assess risk, and can they publish? Put that in your next security questionnaire.
Sources: Anthropic