[whatmodel.sh](../../README.md) / Muse Spark Open # Muse Spark Open Meta · released 2026-10 · 816B params (42B active) · muse-model-license **score 74.4** ±0.9 · #40 of 732 · from 11 benchmark results across 3 sources Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only. ## What it's good at _{● marks Muse Spark Open among every model on that task or test, worst to best.}

task         vs the others            rank  counts  tests
agent tasks  ▁▁▃▆█▄▅▅▄●▂▁      #46 of 313*    1.7%      2
coding       ▁▁▁▃▅▇█▇▅●▃▁      #48 of 469*     14%      1
math         ▁▁▃▅█▆▄█▄●▁▁       #29 of 431     69%      3
web search                  not ranked yet       —      1

agent tasks · #46 of 313 · 1.7% of its score

Trails [GLM-5.2](glm-5-2.md), [GLM-5.3-Flash](glm-5-3-flash.md), [Grok 4.7](grok-4-7.md) here, though they rank lower overall.

test                                result  vs the others in its run  difficulty      counts  why
OSWorld-Verified (maker)             80.4%  █▃·▂▅▂▃▂▅▃●▆ beats 20/27  hard · avg 19%    1.7%  at its level · small run
Humanity's Last Exam with tools (…   61.0%  ▁·▁▂▂▁▄▄▇█●▅ beats 54/62  hard · avg 15%       —  not counted yet: no other model in the …

coding · #48 of 469 · 14% of its score

Beats [GPT-5.5](gpt-5-5.md), [GPT-5.4](gpt-5-4.md), [GPT-5.6 Luna](gpt-5-6-luna.md) here, though they rank higher overall. Trails [Qwen3.8 27B](qwen3-8-27b.md), [GLM-5.2](glm-5-2.md), [Gemini 3.6 Flash](gemini-3-6-flash.md) here, though they rank lower overall.

test                       result  vs the others in its run   difficulty  counts  why
Arena WebDev (Code Arena)  1526.2  ▁▁▃▂▅█▃●▅▅▁▂ beats 81/129  avg 981        14%  sharp

math · #29 of 431 · 69% of its score

Beats [Gemini 3.8 Flash](gemini-3-8-flash.md), [DeepSeek V4.1 Flash](deepseek-v4-1-flash.md), [Grok 4.6](grok-4-6.md) here, though they rank higher overall.

test                       result  vs the others in its run    difficulty            counts  why
FrontierMath tiers 1–3 v2   69.8%  ▂··▁▁▃▄▄█●▄▃ beats 59/86    very hard · avg 1.2%     35%  sharp · at its level
FrontierMath tier 4 v2      31.7%  ▁▂▁▂█●▂▃▂▁▁▁ beats 37/67    very hard · avg 0.1%     34%  sharp · at its level
OTIS mock AIME 2024–2025    98.1%  ▃▃▄▃▃▃▇▇█▃●▅ beats 172/197  very hard · avg 7.2%    1.1%  sharp · maxed out · noisy

web search · 0.0% of its score

test                                result  vs the others in its run  difficulty      counts  why
Humanity's Last Exam with tools (…   61.0%  ▁·▁▂▂▁▄▄▇█●▅ beats 54/62  hard · avg 15%       —  not counted yet: no other model in the …

other tests (8)

test                                result  vs the others in its run    difficulty            counts  why
GPQA Diamond                         88.6%  ▁▁▁▁▂▄▆▄▅▇●▅ beats 166/211  hard · avg 39%          5.3%  sharp
SimpleQA Verified                    56.6%  ▁▁▃▂▃▃▆▆█●▂▅ beats 63/81    very hard · avg 8.0%    3.1%  at its level
Video-MME v2 (with subtitles) (ma…   61.5%  █·········●█ beats 1/2      hard · avg 17%          2.0%  at its level · small run
SimpleVQA with tools (search) (ma…   73.6%  █·········█● beats 2/2      hard · avg 27%          1.9%  at its level · small run
IntPhys 2 (maker)                    68.3%  █···█······● beats 2/2      hard · avg 27%          1.5%  at its level · small run
BabyVision with tools (maker)        80.2%  █······●··██ beats 1/3      hard · avg 30%          0.8%  noisy · small run
ZeroBench main with tools (pass@5…   50.0%  ▄········●·▄ beats 2/3                                 —  not counted yet: no other model in the …
WorldVQA with tools (maker)          49.4%  █··········● beats 1/1                                 —  not counted yet: no other model in the …

## How the score was worked out Each result implies a score on its own, after adjusting for how hard that test is and who ran it. The model's 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 is added up by who ran the tests. [Method](../method.md)

source                                    results  share of the score
Epoch AI                                        5   78% ▇▇▇▇▇▇▇▇▇
LMArena leaderboard dataset                     1   14% ▇▇
Meta vendor-reported benchmark scores f…        5    8% ▇
prior: pull toward the middle                        0% ▇
weighted average                                   = 74.4

_{Licences: Epoch AI ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/))}

Not counted (3)

source                  test                            score  why
Meta vendor-reported …  Humanity's Last Exam with too…  61.0%  too few models have it yet
Meta vendor-reported …  ZeroBench main with tools (pa…  50.0%  too few models have it yet
Meta vendor-reported …  WorldVQA with tools             49.4%  too few models have it yet

## How to run it ### Through an API api (Fireworks AI) · $2.00/M blended · $1.25 in / $4.25 out per 1M tokens · [fireworks.ai](https://fireworks.ai/models/fireworks/muse-spark-open) _{[price source](https://github.com/anomalyco/models.dev/blob/3f893bf02c54a49ebdee0048a812e2ac7f5bd514/providers/fireworks-ai/models/accounts/fireworks/models/muse-spark-open.toml): Fireworks 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. · [context source](https://github.com/anomalyco/models.dev/blob/3f893bf02c54a49ebdee0048a812e2ac7f5bd514/models/meta/muse-spark-open.toml): models.dev entry for Muse Spark Open (pinned): limit.context 1,048,576 tokens, citing Meta's blog, model card and developer docs; Meta's blog states a 1M-token context window. Model ceiling, not a promise by any hosted API. Observed 2026-10-16.} ### llama.cpp (recommended) 1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md) ```sh # macOS / Linux curl -LsSf https://llama.app/install.sh | sh # Windows irm https://llama.app/install.ps1 | iex ``` 2. Run Muse Spark Open (Q4_K_S, needs ~472.25 GB, from [huggingface.co/ggml-org](https://huggingface.co/ggml-org/Muse-Spark-Open-816B-A42B-GGUF/tree/ada96652107984f184027153a77ec250d67ed271)) ```sh llama cli -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q4_K_S ``` 3. Or start a local server and open the address it prints (http://localhost:9931) ```sh llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q4_K_S ``` Other versions: ud-iq1_s (~179.74 GB) `llama cli -hf unsloth/Muse-Spark-Open-816B-A42B-GGUF:UD-IQ1_S` · ud-q2_k_xl (~315.32 GB) `llama cli -hf unsloth/Muse-Spark-Open-816B-A42B-GGUF:UD-Q2_K_XL` · Q2_K (~306.26 GB) `llama cli -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q2_K` · ud-q4_k_xl (~509.6 GB) `llama cli -hf unsloth/Muse-Spark-Open-816B-A42B-GGUF:UD-Q4_K_XL` · Q8_0 (~873.83 GB) `llama cli -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q8_0` ### LM Studio 1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download) 2. Run Muse Spark Open (4bit, needs ~461.7 GB, Mac only, from [huggingface.co/mlx-community](https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-4bit/tree/a030c4bde7045f0352c013cbc7fda64dc2dcbb78)) ```sh lms get https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-4bit -y ``` 3. Then run lms ls to see its name, and lms chat followed by that name Other versions: ud-iq1_s (~179.74 GB) `lms get https://huggingface.co/unsloth/Muse-Spark-Open-816B-A42B-GGUF@UD-IQ1_S -y` · ud-q2_k_xl (~315.32 GB) `lms get https://huggingface.co/unsloth/Muse-Spark-Open-816B-A42B-GGUF@UD-Q2_K_XL -y` · Q2_K (~306.26 GB) `lms get https://huggingface.co/ggml-org/Muse-Spark-Open-816B-A42B-GGUF@Q2_K -y` · 3bit (~359.21 GB, Mac only) `lms get https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-3bit -y` · Q4_K_S (~472.25 GB) `lms get https://huggingface.co/ggml-org/Muse-Spark-Open-816B-A42B-GGUF@Q4_K_S -y` · ud-q4_k_xl (~509.6 GB) `lms get https://huggingface.co/unsloth/Muse-Spark-Open-816B-A42B-GGUF@UD-Q4_K_XL -y` · Q8_0 (~873.83 GB) `lms get https://huggingface.co/ggml-org/Muse-Spark-Open-816B-A42B-GGUF@Q8_0 -y` ### MLX (Mac) 1. Install MLX (Mac). Apple silicon Macs only, text only. ```sh # with uv uv tool install mlx-lm # or pip install mlx-lm ``` 2. Run Muse Spark Open (4bit, needs ~461.7 GB, Mac only, from [huggingface.co/mlx-community](https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-4bit/tree/a030c4bde7045f0352c013cbc7fda64dc2dcbb78)) ```sh mlx_lm.chat --model mlx-community/Muse-Spark-Open-816B-A42B-4bit ``` 3. Or start an OpenAI-compatible server at http://localhost:8080/v1 ```sh mlx_lm.server --model mlx-community/Muse-Spark-Open-816B-A42B-4bit ``` Other versions: 3bit (~359.21 GB, Mac only) `mlx_lm.chat --model mlx-community/Muse-Spark-Open-816B-A42B-3bit` Model repository and documentation: [huggingface.co/meta-models](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) [interactive version](https://whatmodel.sh/model/muse-spark-open) · [JSON](https://whatmodel.sh/api?format=json&all=1)