Meta Muse Spark Open exceeded Nvidia Nemotron by 39 points, but fell behind Chinese models in the 3.5 Super VL，Agent tests

Article

Nvidia had just claimed the top spot among U.S. open-source models,Meta only to 39 points ahead,Agent yet its benchmark score trailed Chinese models

Dongcha Beating AI breaking news: Nvidia's top-ranked U.S. open-source model held the position for only a few hours. Nemotron 3.5 Super VL scored 26.5 points on the Artificial Analysis Intelligence Index, surpassing all previously ranked U.S. open-source models.

But Meta subsequently released Muse Spark Open, which scored 39 points, leading Nvidia by 12.5 points and becoming the new U.S. leader.

However, in the global open-source model rankings,Meta still trails several Chinese models. Xiaomi's MiMo-V2.6-Pro leads by 46 points, while Zhipu's GLM-5.3 is at 45 points, and Muse Spark Open ties with DeepSeek V4.1 Flash at 39 points. It has nearly caught up with the closed-source flagship Muse Spark Meta August released by 1.2, which scored 40 points.

Among the 9 highest-scoring open-source models, Muse Spark Open ranked last in the combined performance of 4 tests, but ranked fourth across the other Agent tests combined. The 6-point gap with Zhipu GLM-5.3 was entirely attributable to the 6 test.

Conversely, in the knowledge-accuracy test,Agent-Omniscience, it led all open-source models with AA points. Its hallucination rate was only 25, the lowest among open-source models scoring more than 24%points.

Muse Spark Open's 30 use efficiency was also solid. It output approximately Token per test on average, 40000 Token fewer than GLM-5.3.

The two models'43%prices were similar, but because Muse consumed fewer API, its average cost per test was only Token, less than half of 0.84 USD but still approximately GLM-5.3 times that of DeepSeek V4.1 Flash.

Original article link 3 https://m.theblockbeats.info/flash/372157

Key points

Nemotron 3.5 Super VL scored 26.5 points on the Artificial Analysis Intelligence Index.
Muse Spark Open scored 39 points, leading Nemotron by 3.5 Super VL 12.5 points.
MiMo-V2.6-Pro scored 46 points,GLM-5.3 scored 45 points, and Muse Spark Open tied with DeepSeek V4.1 Flash at 39 points.
Muse Spark Open ranked last in the combined performance of its Agent tests, but ranked fourth across the other 6 tests combined.
Muse Spark Open scored AA in the 25-Omniscience test, with a hallucination rate of 24%。
Muse Spark Open output approximately 40000 Token per test on average, with an average cost per test of 0.84 USD。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a comparison of model rankings. Muse Spark Open had a higher overall score and performed better in knowledge accuracy and cost, but Agent performed more weakly in the relevant tests.

Why it matters to readers

The results show that overall model scores,Agent capabilities, knowledge accuracy, and usage costs may each have different strengths.

When assessing model rankings, do not focus only on the total score; also consider specific tests,Token consumption, and per-test costs.

Risks and unknowns

The evidence did not provide the original Artificial Analysis rankings page, so the scores, test date, and version cannot be independently verified.
The evidence did not specify the specific items, weights, or samples for the Agent test.
The evidence did not provide the calculation methodology for the hallucination-rate and cost data.

Meta Muse Spark Open exceeded Nvidia Nemotron by 39 points, but fell behind Chinese models in the 3.5 Super VL，Agent tests

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center