As of: 2026-10-15 23:34 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/m/muse-spark-open/

# Muse Spark Open (816B-A42B)

Meta, Muse family. Status: current. Type: reasoning model. Released: 2026-10-15. Training cutoff: 2026-01.

## How to call it

| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face (gated, manual approval) | `meta-models/Muse-Spark-Open-816B-A42B` |  | https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B |
| OpenRouter | `meta/muse-spark-open` |  | https://openrouter.ai/meta/muse-spark-open |
| Baseten, Fireworks AI, Groq, Modal, Nebius, Together AI (hosted APIs named by Meta) |  |  | https://research.meta.ai/blog/muse-spark-open |
| Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry (weights) |  |  | https://research.meta.ai/blog/muse-spark-open |
| Apple MLX (local, 2–4 M5 Ultra Mac Studios) |  |  | https://machinelearning.apple.com/research/mlx-distributed-inference https://github.com/ml-explore/mlx-lm |

## Pricing

- Input: $1.25 per 1M tokens
- Output: $4.25 per 1M tokens
- Unit as published: per 1M tokens (USD) on OpenRouter at launch
- Source: https://openrouter.ai/api/v1/models
- Last checked: 2026-10-15

## Specs

- Input: text, image
- Output: text
- Context window: 1,000,000 tokens
- Max output: not published
- Open weights: yes
- Licence: muse-model-license (limited commercial terms)

## What it has not seen

- Published training cutoff: 2026-01 (January 2026).
- AI events after it on Post-Cutoff: 1,194 (257 major, 42 historic), as of 2026-10-15.
- Already missing at launch: 1,175.
- Briefings of the news since January 2026: https://postcutoff.com/briefings/cutoff-2026-01-s.md (short), https://postcutoff.com/briefings/cutoff-2026-01-m.md (medium), https://postcutoff.com/briefings/cutoff-2026-01.md (full).
- Its card, with five things it has never heard of: https://postcutoff.com/you/muse-spark-open/

## What stands out

- Largest US open-weight model at release: 816B total / 42B active natively multimodal MoE with 1M context; Meta calls it 'the most capable US open-weight model to date' (own evaluation). Source: https://research.meta.ai/blog/muse-spark-open
- Frontier-scale model on desktop machines (first): Runs in 4-bit NVFP4 (~493 GB) across two to four M5 Ultra Mac Studios linked by Thunderbolt 5 RDMA; about 25 tok/s on two and 37 tok/s on four (Apple ML Research). Source: https://machinelearning.apple.com/research/mlx-distributed-inference

## Notes

First Muse model with open weights. Knowledge cutoff inferred from the chat template default cutoff_date 2026-01-04 on Hugging Face (not stated in prose). Reasoning effort low/medium/high/xhigh (default xhigh). Benchmarks not yet independently reproduced.

Open-weight sibling of the hosted Muse Spark models; engines: SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, Unsloth.
</content>
</invoke>

Model page: https://postcutoff.com/m/muse-spark-open/