whatmodel.sh / Muse Spark Open
Muse Spark Open
↗ share
+ contribute
Muse Spark Open is an open-weights language model from Meta, released October 2026. Its published size is 816B parameters (42B active per token) and a 1M-token context window. Licence: muse-model-license.
Catalogue identity notes
Meta blog post 2026-10-15 and the gated (manual approval) repo meta-models/Muse-Spark-Open-816B-A42B@0fb0247: 816B total / 42B active MoE, natively multimodal, weights under the Muse Model License with limited commercial terms. Epoch model_metadata.csv lists 'Muse Spark Open' (Meta AI, 2026-10-15, 'Open weights (restricted use)', low/medium/high/xhigh) with no results yet; not on the LMArena text leaderboard as of dataset revision 7131711 (2026-10-16). Hosted APIs exist (Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI per the blog); the listed price is Fireworks AI's.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
Meta blog post (2026-10-15): 816B total parameters, 42B active (mixture of experts). Hugging Face safetensors metadata for meta-models/Muse-Spark-Open-816B-A42B@0fb0247 counts 822,036,529,664 parameters (BF16 821,906,490,368 + F32 130,039,296), which includes the multimodal encoders.
Score within the LMArena (text) comparison set only.
$ how it compares
The LLMs ranked either side of Muse Spark Open, out of 878 with a score, then more from Meta.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 34 |  Gemini 3.7 Flash | 75±1 | api $1.50/M | ||
| 35 |  GLM-5.3 | 75±1 | api $2.15/M +local | ||
| 36 | Muse Spark Open | 75±1 | api $2.00/M +local | ||
| 37 |  Qwen3.8 Max (0902) | 75±1 | api $3.00/M | ||
| 38 |  DeepSeek V3.2 Speciale | 75±3 | comparison only | ||
| more from Meta | |||||
| 5 | Muse Wonder Full | 83±3 | Muse Code only | ||
| 14 | Muse Wonder | 79±1 | api $8.00/M | ||
| 25 | Muse Spark 1.3 | 76±1 | api $2.00/M | ||
$ what it's good at
agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #49 of 306 * · 2.0% of its score · 2 tests
Trails Claude Sonnet 4.6, Qwen3.8 27B, GLM-5.2 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| OSWorld-Verified lab | 80.4% ✎ | █▃·▂▅▂▃▂▅▃▆▆ beats 20/27 | 2.0% |
| Humanity's Last Exam with tools lab | 61.0% ✎ | ▂▂▂▁▄▂▃▆█▇▃▆ beats 48/55 | — |
math▁▁▃▅█▆▄█▄▃▁▁ #26 of 427 · 81% of its score · 3 tests
Beats DeepSeek V4.1 Flash, Gemini 3.8 Flash, Grok 4.6 here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| FrontierMath tiers 1–3 v2 | 69.8% ✎ | ▂··▁▁▃▄▄█▆▄▃ beats 59/84 | 40% |
| FrontierMath tier 4 v2 | 31.7% ✎ | ▁▂▁▂█▅▂▃▁▁▁▁ beats 37/65 | 40% |
| OTIS mock AIME 2024–2025 | 98.1% ✎ | ▃▃▄▃▃▃▇▇█▃▃▅ beats 172/195 | 1.2% |
web searchnot ranked on this task yet · 1 test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Humanity's Last Exam with tools lab | 61.0% ✎ | ▂▂▂▁▄▂▃▆█▇▃▆ beats 48/55 | — |
other tests8 not in a task · 16% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| GPQA Diamond | 88.6% ✎ | ▁▁▁▁▂▄▆▄▅▇█▄ beats 166/209 | 6.1% |
| SimpleQA Verified | 56.6% ✎ | ▁▁▃▂▃▃▆▆█▄▂▅ beats 62/79 | 3.9% |
| Video-MME v2 (with subtitles) lab | 61.5% ✎ | █·········██ beats 1/2 | 2.3% |
| SimpleVQA with tools (search) lab | 73.6% ✎ | █·········██ beats 2/2 | 2.3% |
| IntPhys 2 lab | 68.3% ✎ | █···█······█ beats 2/2 | 1.8% |
| ZeroBench main with tools (pass@5) lab | 50.0% ✎ | █··········█ beats 1/1 | — |
| BabyVision with tools lab | 80.2% ✎ | █··········█ beats 1/1 | — |
| WorldVQA with tools lab | 49.4% ✎ | █··········█ beats 1/1 | — |
$ where to run it
Through an API
api · Fireworks AI · $2.00/M · $1.25 in / $4.25 out per 1M tokens · 1M context · fireworks.ai ↗ · price sourceFireworks AI serverless standard rate from models.dev (cost input 1.25, output 4.25 USD per 1M tokens; cached input 0.15 excluded). Provider-specific tiers, long-context surcharges, caching and promotions may differ. Fireworks serves text and image input.
llama.cpp · recommended
version
- Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.macOS / Linux
curl -LsSf https://llama.app/install.sh | shWindowsirm https://llama.app/install.ps1 | iex - Run Muse Spark Open · needs ~472.25 GB · from huggingface.co/ggml-orgchat in the terminal
llama cli -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q4_K_S - Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q4_K_S
LM Studio
version
- Install LM Studio · Open the app once before using the lms command.
- Run Muse Spark Open · needs ~461.7 GB, Mac only · from huggingface.co/mlx-community
- Or download it in the terminal
lms get https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-4bit -y - Then run lms ls to see its name, and lms chat followed by that name
MLX (Mac)
version
- Install MLX (Mac) · Apple silicon Macs only, text only.with uv
uv tool install mlx-lmorpip install mlx-lm - Run Muse Spark Open · needs ~461.7 GB, Mac only · from huggingface.co/mlx-communitychat in the terminal
mlx_lm.chat --model mlx-community/Muse-Spark-Open-816B-A42B-4bit - Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/Muse-Spark-Open-816B-A42B-4bit
Weights and documentation: huggingface.co/meta-models ↗
$ how the score was worked out · Epoch AI — internal evaluatio… 91%, Meta vendor-reported benchmar… 8%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 Epoch AI — internal evaluations | 5 | 91% |
 Meta vendor-reported benchmark scores for its own Muse models (Muse Wonder model card, Muse Spark 1.3 evaluation methodology, Muse Spark Open model card) · lab-reported | 4 | 8% |
| prior · pull toward the middle | 0% | |
| weighted average | = 74.9 |
not counted (4)
| Meta vendor-reported benchmark scores for its own Muse models (Muse Wonder model card, Muse Spark 1.3 evaluation methodology, Muse Spark Open model card) | Humanity's Last Exam with tools [correct] | 61.0% | fewer than three models have this test yet, or this source ran it for this model alone |
| Meta vendor-reported benchmark scores for its own Muse models (Muse Wonder model card, Muse Spark 1.3 evaluation methodology, Muse Spark Open model card) | ZeroBench main with tools (pass@5) [correct] | 50.0% | fewer than three models have this test yet, or this source ran it for this model alone |
| Meta vendor-reported benchmark scores for its own Muse models (Muse Wonder model card, Muse Spark 1.3 evaluation methodology, Muse Spark Open model card) | BabyVision with tools [correct] | 80.2% | fewer than three models have this test yet, or this source ran it for this model alone |
| Meta vendor-reported benchmark scores for its own Muse models (Muse Wonder model card, Muse Spark 1.3 evaluation methodology, Muse Spark Open model card) | WorldVQA with tools [correct] | 49.4% | fewer than three models have this test yet, or this source ran it for this model alone |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json