Meta Opens Its Best Model as the Open-Weight Race Tightens

Meta released the weights for Muse Spark Open on October 15, an 816-billion-parameter mixture-of-experts model it calls the most capable US open-weight model to date. The interesting number is not the benchmark score. It is everything wrapped around the release: day-zero support in vLLM and SGLang, a desktop path through Apple MLX, a custom license with commercial limits, and weights parked on Hugging Face, Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry at the same time. Open weights in 2026 are a supply chain, not a download button.

What Meta actually released

Muse Spark Open carries 816 billion total parameters with 42 billion active per token, a 1-million-token context window, and native multimodality across text, image, video and audio, according to Meta AI Research. It is derived from the closed Muse Spark model Meta shipped in August, with extra safety post-training so it holds up as a standalone model without a vendor's system guardrails. Meta also confirmed in the vLLM day-zero post that its 30B Glimmer model was distilled from this same base.

Distribution is unusually wide for a frontier-scale release. Hosted inference is available through Together AI, Fireworks, Baseten, Modal, Groq, Nebius and OpenRouter. Weights are downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Inference support covers SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA silicon, and PyTorch's TorchTitan supports supervised fine-tuning and RL out of the box, per the PyTorch team.

The demo Meta chose to highlight is telling: agentic coding running entirely on-device. Its Muse Code terminal, powered by Muse Spark Open, shipped real iOS changes to Ice Cubes, an open-source Mastodon client, including bug fixes, new features and a full build. The same model drove DaVinci Resolve end to end, reviewing footage frame by frame to pick and crop the strongest moments. That is a workflow claim, not a leaderboard claim, and it is the one builders should hold Meta to.

Three postures for open weights

The Muse Model License allows many developers to use the model for free under "limited commercial terms." That is open weights with a legal fence around them, and it puts Meta in a different camp from its own earlier release: Glimmer 30B shipped under permissive Apache 2.0.

There are now three distinct postures in the market. The first is permissive open weights, where the license barely constrains commercial use. The second is gated open weights like Muse Spark Open, where the file is public but the terms are not. The third is licensed weights, where nothing is published at all. Microsoft made that third choice explicit: it completed pre-training of MAI-Base-2, a 2.8-trillion-parameter MoE with 80 billion active parameters, trained on 16,000 GB200 GPUs over 38 days at an average 21% model FLOPs utilization and 89% goodput, and says it will license those weights to "trusted partners" through Foundry rather than publish them. Microsoft's differentiator is provenance: organic training data, no distillation from other models.

Both companies are also selling the same new currency. Meta ships kernel-level speedups of 1.03x to 3.95x on NVIDIA and AMD GPUs through Helion kernels in SGLang. Microsoft claims its recipe is 4 to 12 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, the strongest open base model, and says it reinvested the savings into a larger model. Efficiency numbers are becoming the headline metric because they map directly to the bill a customer pays, not because anyone can audit them from the outside.

The memory wall moves the deployment math

Sparsity is what makes a model like this thinkable outside a data center. Only 42 billion parameters fire per token, so compute per token looks like a mid-size dense model. But every token still has to read from the whole 816B checkpoint, and that is a memory problem, not a compute problem. Meta's own desktop recipe is two Apple M5 Ultra Mac Studios, 512 GB of unified memory each, joined by Apple's unified memory architecture into roughly a terabyte that behaves like one machine, with 1.2 TB/s of bandwidth fast enough to read the slice each token needs.

That is the real shape of local frontier inference in 2026: a workstation-class deployment, not a laptop one. For anyone evaluating self-hosting, the questions change. Parameter count stops mattering as much as bytes per usable weight after quantization, memory bandwidth at decode time, and whether your inference engine has a day-zero kernel path for your hardware. vLLM serving Meta's NVFP4 checkpoint, with LoRA reinforcement learning already wired up through Verl, is what "day zero" now means in practice.

Safety has also moved. Because the open weights ship without a vendor's system guardrails, Meta hardened the weights themselves against misuse as a standalone model. Every builder who deploys it inherits that responsibility for their own users.

What builders should do next

Decide on license before you decide on benchmarks. If your product is commercial, read the terms first, because a model you cannot legally ship inside your product is not a model you adopted.

Start hosted, self-host later. Groq, Together, Fireworks, Baseten, Modal, Nebius and OpenRouter all serve this model today, which is the cheapest way to measure it against your actual workload. Move to your own hardware only when data control or volume justifies a terabyte of memory and the ops load that comes with it.

Budget per task, not per token. Agent loops burn tokens on retries and tool calls, so the metric that matters is the cost of finishing a job. That is the same shift we covered in today's look at Jev's adoption data, where routing and model choice became a measured line item rather than a preference.

Assume your guardrails are yours. An open-weight frontier model with no built-in system guardrails means policy, logging and refusal behavior live in your application layer. If you skip that work, you inherit the risk instead of the prize.

Watch the license drift. When the same weights appear on Foundry, SageMaker and Model Garden under a custom license, "open" is increasingly shorthand for "downloadable under terms." The stack around the model now decides what you can build, and this week two of the largest labs made that explicit.