Muse Spark Open: Meta Releases 816B Open-Weight Coding Model

What Meta shipped

On October 15, Meta released downloadable weights for Muse Spark Open. The company calls it the most capable open-weight model a US lab has released so far.24 It is a sparse mixture-of-experts system with 816 billion total parameters. About 42 billion of them are active for any given token, and it can hold up to 1 million tokens of context.311 It accepts text, images, audio and video, and it answers in text.1116

This is not a brand-new model. It is an open version of Muse Spark, the frontier model Meta released in August, with extra safety training on top. That training is meant to make it harder to misuse when it runs without the guardrails Meta applies to its own hosted service.1213 PyTorch's announcement says it shares a base model with Muse Spark 1.3.13 One analysis says the real change is ownership: Muse Spark was already available through Meta's own products, and now developers can download the weights and run them themselves.6

Meta launched with a lot of partners already lined up. Hosted access was available on day one through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.2 The weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.2 On launch day it worked with SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth, with tuning for AMD, Apple, Intel and Nvidia chips.5

Under the hood

The technical details are consistent across outlets. The model has 62 layers and a hidden dimension of 8,192. Each layer holds 256 experts, and a router picks 8 of them per token.1116 Those experts account for about 97% of the parameters. That is why the model computes each token about as cheaply as a 42-billion-parameter model, but still needs memory for the full 816 billion.11 A separate 1.8-billion-parameter Perception Encoder processes visual input.16

Baseten described two design choices that matter for coding agents. One is a compressed expert block that cuts traffic between GPUs. The other mixes 46 local-attention layers with 16 global-attention layers. Baseten says this shrinks the short-term memory each conversation needs (the KV cache) from 31 GiB to 8.4 GiB per sequence at 128,000 tokens of context. Meta also shipped Helion GPU kernels, which are building blocks written in PyTorch's new kernel language. It reports speedups of 1.03 to 3.95 times on individual operations and a 3% overall throughput gain on AMD's MI350.13

The exact figures vary a little between sources, mostly at the margins. Baseten puts the active parameter count at 42.7 billion. Hugging Face measures the BF16 checkpoint at about 1.5 TB, while Meta's own guide says about 1.6 TB.611 None of these gaps changes the overall picture.

The desktop claim, and its fine print

The most talked-about angle is Apple hardware. Meta says the model runs on two M5 Ultra Mac Studios linked with Apple's MLX framework. Each machine holds half of every layer, and only small intermediate results pass between them.14 Each machine can be configured with up to 512GB of unified memory, so the pair has about 1TB combined. TechCrunch framed the launch as a joint effort between Meta and Apple.17 The New Stack carried a social-media post crediting work with Apple.

"Desktop" should be read loosely. TechCrunch, citing Engadget, says local use needs at least 1TB of RAM.17 On servers, the model card asks for at least 700GB of GPU memory for the 4-bit NVFP4 version and 2TB for full BF16 weights.4 The NVFP4 build fits on four Blackwell Ultra GPUs or eight H200s, and community 2-bit versions cut it to about 300GB.5 Hugging Face tested a 3-bit version on a single 512GB M3 Ultra. It used about 344GB and generated roughly 21 tokens per second.11 That is usable, but it is a workstation-class setup, not a laptop.

The practical reading is that most teams will use this model through a hosting provider. Launch pricing on Baseten and Nebius was $1.25 per million input tokens and $4.25 per million output tokens.16 One guide's advice was simple: use a host unless data-residency rules prevent it.6

Benchmarks: strong at agents, weaker at coding

Meta pitches Muse Spark Open as a coding model, but its own benchmark numbers tell a mixed story. On MCP-Atlas, a tool-use benchmark, it scores 89.2, ahead of Kimi K3 (82.3), GLM 5.3 (84.2) and Qwen3.8 (84.5).16 On Terminal-Bench 2.1, a test of command-line coding work, it scores 76.4 and falls behind all three. Kimi K3 scores 85.0 there.16 On the overall Artificial Analysis Intelligence Index, Meta's table puts it at 39, below Kimi K3's 44 and GLM 5.3's 45.16

The New Stack reaches a similar verdict. Against the other US open models, Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra, Meta's numbers show it winning across the board, often by more than 20 points. Against the leading Chinese models, it leads on agent tasks, roughly ties on SciCode and trails on coding. MarkTechPost puts it mid-pack among large open models on coding.16

Baseten compared it with closed models from the big labs. It says that in routine agent loops (running commands, reading output, fixing errors) the model should feel close to the leading closed systems. On large multi-file changes, it expects Muse Spark Open to need more guidance or retries than Anthropic's Claude Opus 5 or OpenAI's GPT-5.6 Terra. It says the model holds up well against xAI's Grok 4.5 and Google's Gemini 3.6 Flash. Baseten sells access to the model, so this assessment is not neutral, but it is more measured than Meta's headline claim.

All of these scores come from Meta or its launch partners. No independent tests have confirmed them yet, and Meta's full evaluation report is still to come.5

My reading: "most capable US open-weight model" is probably accurate, mainly because the US competition is weak. "Best open model" would not be accurate. On coding, which Meta emphasizes most, the leading open models are still Chinese.

The licence is open, but with conditions

Muse Spark Open uses a new Muse Model License, not the Llama licences behind Meta's earlier open releases.8 It allows free use, modification, fine-tuning, distillation and redistribution. Every product built on it must display "Built with Muse Spark Open" somewhere visible.4 Hosting providers that sell access to the model and earn more than $20 million in any twelve-month period need a separate agreement with Meta. Internal use, research and products that embed the model are exempt.6

Coverage splits on how to judge this. The New Stack says that, as with some Chinese releases, the weights are open but the licence is not fully open. Digital Applied argues the terms are friendlier to most businesses than the revenue clause suggests at first.6 Meta describes them as "limited commercial terms" that let many developers use the model for free.2 Both views are fair. For a typical developer or startup, the only real obligation is the attribution line. The revenue clause mainly targets large hosting providers that might otherwise resell Meta's work at scale.

Why it matters: the open-weight race with China

The release matters most as a strategic move. Several outlets present it as Meta's response to the Chinese open models many developers already use.5 Shattered.io argues that the real story is distribution. Meta lined up seven hosting partners, same-day chip support and inference-library integration within a day, rather than simply posting weights.8 Its comparison table puts the model in the same size range as other recent large open releases, including Mistral's 675-billion-parameter Large 4 and Tencent's 770-billion-parameter Hy4.8 It argues that the 1-million-token context combined with native audio and video input is what sets Muse Spark Open apart.8

The launch also fits a broader shift at Meta. The company has moved toward closed flagship models since April, and this release follows Muse Wonder, its first frontier model.5 Muse Glimmer, a small dense model in the same family released in August, covers the on-device end.11 The pattern looks deliberate. Meta keeps its newest flagship closed and releases a slightly older, safety-hardened version with open weights to win over developers.

Safety is part of the pitch. Meta says the model passed the release thresholds of its Superintelligence Scaling Framework, which covers content safety, agentic risk, privacy and preparedness.4 Technology.org cites a study finding that Chinese labs publish safety tests for only 3.6% of their model releases.5 Meta clearly hopes buyers will see published safety testing as a reason to pick a US model.

Meta is not claiming to have beaten every open model, and its own numbers would not support that claim. It is offering US developers and companies a domestically built model that is competitive on agent work and easy to get, from a company that publishes safety evaluations. Independent benchmarks will show whether that is enough to pull developers away from Kimi, GLM and Qwen.