NVIDIA’s Nemotron 3.5 Score Doubles, Yet The Gap Remains

🔍 Read the full analysis: NVIDIA’s Nemotron 3.5 Score Doubles, Yet The Gap Remains on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released two open-weight reasoning models on October 15, and independent evaluator Artificial Analysis scored the flagship Nemotron 3.5 Super VL at 26 on its Intelligence Index, twice the 13 recorded for Nemotron 3 Super seven months earlier. The result strengthens NVIDIA’s position among US models with permissive licenses, while higher-scoring open-weight models from China and the United Arab Emirates show that the wider rankings remain led by other countries.

Super VL has 120 billion total parameters, of which 12 billion are active for each token, and accepts text and images. Its OpenMDW-1.1 license allows commercial use, according to the source report. Artificial Analysis recorded a score of 26, or 26.5 before rounding. The same evaluation lists its predecessor at 13. Super VL used about 49,000 output tokens per index task, compared with about 114,000 for the previous model. The report puts cost per task at $0.25, down from $1.64.

The second release, Nemotron 3.5 Bolt 74B-A9B, scored 21. It has 74 billion total parameters and 9 billion active parameters, compared with 550 billion and 55 billion, respectively, for NVIDIA’s Nemotron 3 Ultra, which scored 23. Bolt ranked fourth among models below 100 billion total parameters in Artificial Analysis’s listing and was the highest-scoring US model in that size group, according to the report. It has no public API yet, and Artificial Analysis had not listed its license or measured a task price.

Artificial Analysis also reported a hardware and serving result for Super VL. Its NVFP4 version scored within 0.1 points of full-precision BF16 weights in the evaluator’s comparison; no individual evaluation differed by more than 2.6 points. On a pre-release DeepInfra endpoint, the evaluator measured about 630 output tokens per second with 100,000-token prompts and 1.8 seconds to the first token. The report says the median speed across six providers on the model page was 196 tokens per second, illustrating that serving performance varies by provider.

A Stronger US Midrange Model

The release offers developers a model that combines a higher evaluation score with lower reported task cost than its predecessor. For organizations considering self-hosting, the report’s NVFP4 results also suggest that a lower-memory format can retain a score close to the BF16 version in these tests. Those figures may help teams compare options, but actual speed and cost depend on deployment and provider; the reported index-task cost is not a universal operating cost.

The rankings put the improvement in perspective. Super VL is described as the leading US open-weight model with a permissive license, but the report places it 18th among all listed open-weight models and 11th among permissively licensed models. The best permissively licensed model listed, Xiaomi’s MiMo-V2.6-Pro, scored 46, twenty points above Super VL. The gap matters for buyers who weigh model capability alongside licensing and the practical ability to use a model commercially.

How the Rankings Compare

Artificial Analysis’s Intelligence Index is the basis for the scores cited in the source report. The reported leaderboard includes models with different sizes, licenses and costs, so a score alone does not capture every deployment consideration. The report says nine of the ten highest-scoring open-weight models it counted came from Chinese labs. The remaining model in that group was Meta’s Muse Spark Open, which scored 39.

Muse Spark Open was released on the same day as NVIDIA’s models. The source describes it as a much larger model, with 816 billion total parameters and 42 billion active, and reports a cost of $0.84 per index task. It uses the Muse Model License, which the source says has commercial restrictions. That license distinction affects how the models may be used; the source does not provide the full license terms here.

Among other models cited, the UAE’s K2 Horizon 375B A23B scored 31 and carries an Apache 2.0 license, according to the report. NVIDIA’s Nemotron 3 Ultra scored 23 and cost $0.60 per index task. Super VL’s score and reported cost compare favorably with that earlier NVIDIA flagship, although the models differ in size and evaluation results.

“The honest framing.”

— Thorsten Meyer, source report

Limits of the Reported Results

The source material does not specify the full evaluation methodology or establish how closely Intelligence Index scores predict results for particular workloads. The reported cost per task is tied to the evaluator’s index task and should not be read as a standard price for all users. Provider speeds also vary: the cited 630-token-per-second measurement came from a pre-release DeepInfra endpoint, while the reported six-provider median was lower.

Bolt’s commercial availability and economics remain unsettled in the source material. It had no public API at the time of the report, and Artificial Analysis had not listed its license or measured its cost per task. The source also cuts off while describing the Muse Model License, so its detailed restrictions cannot be established from the available material.

API Access and Leaderboard Changes

The next practical milestone for Bolt is a public deployment option and published licensing and pricing information. Until those appear, potential users cannot compare its hosted cost with Super VL or verify how the model’s long outputs affect the cost of a particular answer. The source report also points to self-hosting as an option for NVIDIA’s releases, though performance will depend on available hardware and implementation.

Artificial Analysis’s listings may change as it evaluates additional models and providers. For now, the reported scores establish a clear improvement over NVIDIA’s earlier models, but they do not show the US overtaking the higher-ranked open-weight models cited in the report. Updated evaluations, access terms and real-world serving measurements will show whether that position changes.

Key Questions

What score did Nemotron 3.5 Super VL receive?

Artificial Analysis gave it 26 on the Intelligence Index, or 26.5 before rounding, according to the source report.

How does Super VL compare with its predecessor?

Nemotron 3 Super scored 13 in the comparison cited by the report. Super VL scored 26 and used about 49,000 output tokens per index task, versus about 114,000 for its predecessor.

Is Super VL the highest-scoring open-weight model?

No. The report places it 18th among listed open-weight models. It describes Super VL as the leading US model with a permissive license, while Xiaomi’s MiMo-V2.6-Pro scored 46.

Can users access Bolt through an API?

The source report says Bolt had no public API at the time of publication. Artificial Analysis had not listed its license or measured a cost per index task.

Source: ThorstenMeyerAI.com