Muse Spark Open
Meta's open-weight model, derived from Muse Spark with extra safety training, which Meta calls the most capable open-weight model from a US lab. The 816B-A42B weights are on Hugging Face and the model is also offered by major cloud hosts.
MetaLanguage modelReleased October 2026Data checked October 18, 2026
This model is not in the current people's rankings: it is either new and still collecting votes, or ranked below the top 12. See the AI leaderboard for the models people rate highest.
On paper
Fixed-test rankings from the benchmarks tab: Speed #7, Reasoning #13.
Strengths and weaknesses
- Open weights under Meta's own Muse model license
- Artificial Analysis Intelligence Index of 38.9, on par with GPT-6 Luna and Mistral Large 4
- Fast on Artificial Analysis (131 tokens per second)
- Very large, so self-hosting needs serious GPU hardware
- Not yet on the LMArena text leaderboard, and weak at terminal-based agent tasks (8.1% on Terminal-Bench 4.0 as run by Artificial Analysis)
Best for
- Self-hosting a US-built open model
- Private deployments on your own cloud
Every score we have
Each bar compares this model with the best result any model on this page has on the same test.
- LMArena WebDev (coding)1,511.3Arena score 1511.3 (95% CI 1491.4 to 1531.2); 991 votes; entry "Muse Spark Open (Max)"; position 55 of 144 on the full board
- Artificial Analysis Intelligence Index38.938.9 on the Artificial Analysis Intelligence Index v4.3.2 (AA shows 39), setting "Muse Spark Open (xhigh)"
- Humanity's Last Exam (Artificial Analysis)43.9%43.9% on Humanity's Last Exam (AA, 2,158 text-only questions), setting "Muse Spark Open (xhigh)"
- GPQA Diamond, graduate-level science89.0%89.0% on GPQA Diamond (AA; no longer part of the AA index since v4.2), setting "Muse Spark Open (xhigh)"
- Terminal-Bench 4.0 (Artificial Analysis run)8.1%8.1% on Terminal-Bench 4.0 as run by AA (mini-swe-agent harness), setting "Muse Spark Open (xhigh)"
- Long-context reasoning (AA-LCR)78.3%78.3% on AA-LCR (long-context reasoning), setting "Muse Spark Open (xhigh)"
- MMMU-Pro, image understanding80.9%80.9% on MMMU-Pro (image understanding, AA), setting "Muse Spark Open (xhigh)"
- AA-Omniscience, accuracy minus hallucination24.724.7 on the AA-Omniscience Index (knowledge accuracy minus hallucination, range -100 to 100), setting "Muse Spark Open (xhigh)"
- Output speed, tokens per second131 t/s131 output tokens per second (AA median), row "Muse Spark Open (xhigh)"
- Cost per Artificial Analysis test task$0.84$0.84 to run one AA Intelligence Index task, row "Muse Spark Open (xhigh)"
- Artificial Analysis Coding Agent Index41.241.2 on the AA Coding Agent Index v1.5 (DeepSWE v1.1, Terminal-Bench 4.0, SWE-Atlas-QnA), best listed setup "Muse Code - Muse Spark Open (xhigh)"
Compare with
- Claude Opus 5.5Anthropic's flagship for long-running agentic coding and knowledge work, and the model Anthropic suggests starting with for most jobs.
- Claude Sonnet 5.5Anthropic's mid-priced model, which Anthropic describes as its best mix of speed and intelligence.
- Claude Fable 5.1Anthropic's top tier for the hardest reasoning and long-horizon agent work, which Anthropic recommends when Opus 5.5 still falls short.
- Claude Haiku 5.5Anthropic's smallest and fastest current model, built for high-volume, latency-sensitive jobs like classification, extraction and routing.