Model reference · verified 2026-10-17
Meta published the weights for Muse Spark Open on 15 October 2026: 816B total parameters with 42B active per token, native image, audio and video input, and a 1,048,576-token window. It is the most capable open-weight model a US lab has released, and it is also four hundred and ninety-three gigabytes before the KV cache. Both halves of that sentence are the story.
Weights under the Muse Model License On OpenRouter → Meta's model page →
A sparse mixture-of-experts transformer derived from the Muse Spark line, with additional safety post-training for standalone release. Text, image, audio and video in; text out. Meta positions it for agentic coding and long-horizon software work rather than chat.
1.25 / 4.25 per million tokens on OpenRouter, with cached input at 0.15. Self-hosting is a hardware purchase: about 493 GB in NVFP4, or 1.65 TB of BF16 weights, before the KV cache.
Not the frontier score — it is the fact that a model in this class leaves a lab's API at all. Meta says two M5 Ultra Mac Studios can hold it, and Apple measured more than 36 tokens per second on four of them. That is a deployment story, not a demo.
The weights are gated, the licence is not OSI, and the cheapest local build in the first 48 hours used two DGX Sparks and a 194.6 GiB quantization. The model is open; the hardware is not free.
Meta's launch table, read 2026-10-17 Rival numbers are the same table's, on the same row — not a harness we ran
| Category | Benchmark | Muse Spark Open | For scale |
|---|---|---|---|
| Agentic | JobBench | 61.8 | Inkling 31.1 · Nemotron 3 Ultra 27.2 |
| MCP-Atlas | 89.2 | Inkling 76.0 · Nemotron 3 Ultra 63.1 | |
| GDPval-AA v2 | 1401 | Qwen3.8 1613 · GLM 5.3 1653 — the only row where Muse trails the listed rivals | |
| Knowledge & Reasoning | HLE-Diamond · no tools | 26.1 | Inkling 11.2 · Nemotron 3 Ultra 8.9 |
| SciCode | 58.5 | GLM 5.3 59.0 · Mistral Large 4 54.2 | |
| AA-Omniscience | 25 | GLM 5.3 14 · Mistral Large 4 −5; a negative score means more wrong answers than right ones | |
| Coding | Terminal-Bench 2.1 · Terminus 2 agent | 76.4 | GLM 5.3 83.9 · Qwen3.8 82.0 |
| DeepSWE 1.1 · best available / self-reported harness | 60 | GLM 5.3 67 · Mistral Large 4 62 | |
| Multimodal (with tools) | ZeroBench | 50.0 | Inkling 14.0; the other four columns say not supported |
| BabyVision | 80.2 | Inkling 34.2; the rest not supported | |
| SimpleVQA | 73.6 | Inkling 52.3; the rest not supported |
Every row, and what each one does not measure →
| Route | What it is | Memory | Source |
|---|---|---|---|
| BF16 reference | The full-precision checkpoint, served on eight GPUs such as 8× B300. | ≈1.65 TB of weights; budget roughly 2 TB of VRAM for KV cache, activations and runtime overhead | Meta cookbook |
| NVFP4 checkpoint | Most weights in 4-bit with some layers kept in BF16; serves on a single 4-GPU NVIDIA Blackwell node. | ≈537 GB of weights | Meta cookbook |
| MLX on Mac Studio | The desktop route: pipeline or tensor parallelism over two to four M5 Ultra machines. | ≈493 GB in NVFP4; two Macs give 1 TB of unified memory, four give 2 TB | Apple ML Research |
| Two DGX Sparks | A community recipe at IQ1_M on two GB10 nodes over llama.cpp RPC. | 194.6 GiB of weights, about 97 GiB per node | Author's repository |
| Hosted | Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter or Together AI — no local memory at all. | None locally; billed per token, $1.25/$4.25 on OpenRouter | Meta cookbook |
Apple's four-Mac Studio run: 36+ tokens/s at 8,192 context (2026-10-16) The arithmetic behind each route →
Repository
Official recipes for serving and post-training Muse Spark Open: BF16 and NVFP4 checkpoints, SGLang and vLLM, hosted providers, prompting and fine-tuning.
Repository
An MLX layer that runs a 346 GB quantized Muse Spark Open checkpoint on a 128 GB Mac by caching 25% of experts in memory and streaming the rest from the internal SSD.
Repository
A reproducible two-node DGX Spark recipe for Muse Spark Open using llama.cpp RPC and a 194.6 GiB UD-IQ1_M quantization.
Three of the four builds we hold; the fourth is on the builds page. Stars read 2026-10-17. A three-day-old model has a short list — the point of printing it now is that the list will look different in a week.
Hacker News · 914 points
The main Hacker News discussion of the release: parameter count, memory footprint, licence terms and what 'open weights' should mean.
Hacker News · 362 points
Discussion of Apple's MLX and RDMA benchmark: pipeline versus tensor parallelism, NVFP4 memory footprint, and the four-Mac Studio setup.
Hacker News · 72 points
An earlier submission of Meta's launch post, with a short comment thread.
Muse Spark Open arrived with a launch post and a cookbook, which is the tidy version. The next one may arrive as a repository at midnight. Leave an email and we will send one note when a model appears, with the specs read from the catalogue.
One email per model. No newsletter, no forwarding, unsubscribe in one click.
Meta's cookbook → The gated weights The full reference What running it costs