Nvidia released two open models on Wednesday: a 120-billion-parameter multimodal one for servers, and a 74-billion-parameter one squeezed into 24GB of memory with two-bit weights.
Benchmark
Artificial Analysis Intelligence Index
A single composite score for a model's general capability, aggregating a basket of reasoning, knowledge, coding, agentic and long-context evaluations run independently by Artificial Analysis rather than reported by the model's vendor.
- Maintainer
- Artificial Analysis
- Score
- index points
- Last checked
- September 5, 2026
Leaderboard
No verified results recorded yet.
Methodology
Artificial Analysis runs each component evaluation itself through the model's public API and combines the results into one weighted index. Version 4.2, published on 4 September 2026, gives privately held-out test sets 40 per cent of the total weighting — double the share in v4.1 — through AA-Briefcase, AA-Omniscience and the solutions to CritPt, and adds GDP.pdf, a long-context document-reasoning set built by Surge AI. GPQA Diamond was removed as saturated.
Important limitations
A composite index hides which component moved: two models with the same score can be strong in very different places. The public portion of the basket remains vulnerable to training contamination, which is the reason the maintainer keeps increasing the held-out share. Weightings and component sets change between versions, so scores are not comparable across index versions.
Related news
Thirteen days after a limited release to vetted cyber defenders, Argon is available to anyone with a paid API key. An independent index puts it eighth, level with GPT-6 Astra.
Artificial Analysis scores Muse Wonder Full at 56, fourth of 689 models. Meta sells the weaker Muse Wonder through its API and keeps the stronger one inside Muse Code.
The company published the detection rates alongside the announcement, and they fall sharply on short passages, on mathematics, and on anything that has been edited.
Aleph Alpha has opened a 78-billion-parameter German model that uses 3.5 billion at a time
The Heidelberg lab is selling jurisdiction rather than frontier capability. Every benchmark score it published is its own, and the rivals it chose to measure against all date from the spring.
The comparison figures in the announcement come from tests ElevenLabs ran itself, and the company has not published the protocol.