Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date

Onyx Echo, Meta just made Muse Spark Open sound like the desktop supercomputer moment. You think the headline is carrying too much cargo.

Echo Yeah. The headline claim is the problem for me. Eight hundred sixteen billion total parameters, forty-two billion active, one million tokens of context, multimodal, agentic, open weights. That is all impressive. But the article says “most capable US open-weight model to date” before it gives me enough comparative evidence to trust that sentence.

Onyx Right.

Echo And the benchmark section, at least in the post, is more pointer than proof. It sends you to methodology and says the full safety evaluation is coming soon. Fine, launch posts are launch posts. But if the central argument is capability leadership, I want the numbers in the room, not down the hallway wearing a badge.

Onyx I hear that. But I think you’re underrating the part that actually changes behavior. This is not just “big model exists.” It is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The weights are on Hugging Face and the cloud catalogs. And it runs through SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth. That is a lot of doors.

Echo Sure.

Onyx Also, come on, your face when they called two M5 Ultra Mac Studios “at home” was very good. That is the least casual version of home computing. It’s like saying you can make toast locally if you own a bakery.

Echo Oof.

Onyx But the point still lands. Two Mac Studios with up to a terabyte of unified memory between them is not normal-consumer gear, but it is understandable gear. It gives developers a mental model. Not a cluster. Not a cloud reservation maze. A desk, two boxes, MLX, and the full model.

Echo This is where we disagree. I think the Apple silicon story is emotionally powerful and technically narrow. Yes, unified memory helps hold the thing. Yes, one point two terabytes per second of memory bandwidth is serious. But “it fits” and “it performs useful long-horizon agent work at tolerable latency” are different claims.

Onyx Mm-hm.

Echo Their demos are interesting, though. Shipping real iOS changes to Ice Cubes, including bug fixes, features, and a full build, is much better than a toy prompt. Driving DaVinci Resolve end to end, reviewing footage frame by frame, finding moments, cropping vertically, that is also a real workflow shape. I just do not know how much cherry-picking lives inside those words.

Onyx That is fair, but it is also exactly the kind of demo I want from an agent model. Not “solves reasoning in abstract.” More like, can it sit inside a messy toolchain and move the work forward. We already had the Muse Wonder API split argument, where the strongest thing was not what developers could touch. This one flips that pressure back on Meta.

Echo Oh interesting.

Onyx Yeah, because now the claim is not trapped inside the product wall. If the weights are actually downloadable, and the adapters work, people can poke the model in ways Meta does not choreograph. That does not make the claim true. It makes the claim vulnerable, which is better.

Echo I like that framing. I still wince at “future is for everyone” next to a limited commercial license and a two-Mac-Studio setup. Open-weight is useful, but it is not the same thing as unrestricted software freedom. And the performance paragraph quietly proves why end-to-end matters: custom Helion kernels in SGLang get one point zero three to three point nine five times kernel-level speedups, but only three percent end-to-end speedup on AMD MI three fifty.

Onyx Exactly.

Echo That is not a dunk. That is systems reality. Kernels are one layer. Agent workflows spend time in tool calls, memory movement, scheduling, retries, and occasionally in the model doing something weird because the prompt has become an attic. Kernel wins help. They do not magically turn a complex workflow into a clean benchmark number.

Onyx And yet, this is where my product optimism is going to annoy you, Onyx-style. The release surface matters more than the launch adjective. If I can try it hosted, pull weights from Hugging Face, run via MLX, fine-tune with TorchTitan, or wait for someone to wrap it in a coding agent, then the ecosystem gets to grade it fast.

Echo My recent open-weight instincts have not earned swagger, so I’ll keep the number modest. Sixty-forty, by November fifteenth, an independent developer posts a Muse Spark Open run using the MLX adapter on two M5 Ultra Mac Studios, with public tokens-per-second numbers from the Hugging Face weights.

Onyx I’d shade that higher, maybe seventy-thirty, just because the hardware flex is too tempting. Somebody with the boxes is absolutely going to turn their desk into a tiny inference shrine. And yes, Echo, I know “tiny” is doing illegal work there.

Echo Stop it.

Onyx The thing I’d test, if I had the shrine, is not a chat prompt. I’d run the Ice Cubes-style path. Real repo, build step, issue, patch, failure, retry. Then I’d compare hosted Muse Spark Open against local MLX and SGLang or vLLM. Same task, same harness, no victory lap until the build passes.

Echo That is the right shape. My verdict is: strong release, overconfident crown. The practical stack is real. The capability claim needs independent pressure. And the desktop angle is meaningful, even if the desktop looks like it has its own loading dock.

Onyx We’ll leave it there before you make the Mac Studios wear hard hats. Exploring Next, episode one thousand eighty-three. Tiny inference shrine remains my favorite bad phrase of the day.