Muse Spark Open
Meta·
Overview
Open-weight release from Meta Superintelligence Labs: an 816B mixture-of-experts model activating ~42B parameters per token, with a 1M-token context and text, image, video and audio input (text output). It is a separate model from the API-only Muse Spark 1.x line and from Muse Wonder, which Meta launched earlier the same week and keeps closed; Meta states it will generally not open-source its most capable models. Meta says the open checkpoint received additional safety post-training because it runs without Meta's system-level guardrails; the safety evaluation behind the release had not been published at launch. Weights are on Hugging Face under the Muse Model License, which permits use, modification, fine-tuning, distillation and redistribution but requires a prominent "Built with Muse Spark Open" notice. Meta's model card sets a floor of ~700 GB of accelerator memory for the NVFP4 checkpoint and ~2 TB for BF16 (e.g. four Blackwell Ultra GPUs or eight H200s for NVFP4); community GGUF quantizations exist down to 2-bit.
Capabilities and innovations
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 89% | vendor GPQA Diamond 89.0 from Meta's model card (Muse Spark Open at xhigh reasoning). Vendor self-report; no independent run located as of 2026-10-17. The card was read via launch coverage quoting it, since huggingface.co was not reachable from this environment. |
| Terminal-Bench 2.x | 76.4% | vendor Terminal-Bench 2.1 76.4 from Meta's model card, Terminus 2 agent at xhigh. Terminus 2 is the harness the Vals AI reference values in this field use, but this is Meta's own run, not a Vals AI or official-leaderboard entry. Replace with the Vals AI figure once it exists. Do not compare against terminal_bench_3 or terminal_bench_4. |
| DeepSWE v1.1 | 60% | vendor DeepSWE v1.1 60 from Meta's model card, labelled 'preferred harness': Meta's choice of agent, not the fixed mini-swe-agent harness DeepSWE scores under. Same caveat as the Muse Spark 1.3 entry, and stronger: here the card says outright that the harness is not the reference one. Vendor self-report over its own model; the competitor rows in the same table (Kimi K3 68, among others) are Meta measuring other labs' models and are not entered anywhere. Datacurve's board was not reachable from this environment, so the value is not cross-checked. |
- Humanity’s Last Exam (no tools): not reported by the vendor — Meta's model card reports HLE with tools (61.0) and HLE-Diamond without tools (26.1). Neither is the full no-tools HLE used here, so neither goes in the field. A 43.9% no-tools figure attributed to Artificial Analysis circulates in launch coverage; artificialanalysis.ai was not reachable from this environment to confirm it, and 43.9 is also exactly the value this dataset carries for Muse Spark 1.2, so it is not entered until checked at the source.
- SWE-bench Pro: not reported by the vendor — Meta reports DeepSWE v1.1 (recorded in deepswe_v1_1) and SWE-Atlas QnA (58, no field) instead of SWE-bench Pro, as for Muse Spark 1.2 and 1.3.
- MMLU-Pro: independent evaluation pending — Not in Meta's model card; no independent MMLU-Pro run located as of 2026-10-17.
- Terminal-Bench 4.0: independent evaluation pending — One launch write-up (kingy.ai) reports Terminal-Bench 4.0 8.1 at xhigh without naming who measured it; not on Meta's model card, and the official tbench.ai 4.0 leaderboard was not reachable from this environment. Not entered until a primary source confirms it.
- MMMU-Pro: not reported by the vendor — The model card's multimodal rows are ZeroBench, BabyVision, SimpleVQA and WorldVQA; none is tracked here.
Architecture and hardware
- Parameters
- 816B total, 42B active per token (MoE)
- Estimated VRAM at Q4
- ~470 GB, Frontier class
Links
More from Meta
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.