Nemotron 3.5 Super VL And The Appeal Of Sub-Second Latency

🔍 Read the full analysis: Nemotron 3.5 Super VL And The Appeal Of Sub-Second Latency on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, presenting a fast, commercially usable open-weight model for text and image input. Artificial Analysis gave it an Intelligence Index score of 26.5; the accompanying measurements point to low latency and lower benchmark-task costs than its predecessor, but also show that several competing models score higher for comparable spending.

The model has 120 billion total parameters, with 12 billion active in its mixture-of-experts design, and a reported one-million-token context window. NVIDIA released weights in NVFP4, a four-bit format. The stated license, OpenMDW-1.1, allows commercial use. Artificial Analysis lists API prices of $0.20 per million input tokens and $0.575 per million output tokens, with cached input at $0.055 per million tokens.

Artificial Analysis scored the model at 26.5 on its Intelligence Index, roughly double Nemotron 3 Super’s score of 13. It ranked first among 67 open-weight models in its 40-billion-to-150-billion-parameter size class, according to the source account; across open-weight models overall, it ranked about 18th. The score is a benchmark result, not a measure of performance on every production workload.

Speed figures vary by serving provider. A pre-release DeepInfra endpoint produced about 630 output tokens per second with 100,000-token prompts. Across six providers, Artificial Analysis reported median output speeds of 196 tokens per second on its standard prompt and 307 with 100,000-token prompts. Median time to first token was 0.89 seconds. In reasoning mode, however, the source reports about 11 seconds before the first answer token and 13.6 seconds for a 500-token response.

Low Latency Has a Trade-Off

For teams operating document-heavy pipelines or background agents, Super VL’s combination of a long context window, fast generation and low listed input prices could make it useful. The reported median first-token time is under one second, and the source estimates an Artificial Analysis Index task costs about $0.25. That is substantially below the reported $1.64 per task for Nemotron 3 Super.

But the result is not a simple case for choosing Super VL as the best-value model. The source’s comparisons list MiMo-V2.6-Pro at a score of 46 for $0.13 per task, and Inkling Small at 26 for $0.09. Those figures suggest Super VL’s strongest case is not raw benchmark performance per dollar. Rather, it combines speed, commercial-use licensing and the option to run the weights on NVIDIA hardware. Teams should weigh those attributes against their own latency, legal and infrastructure requirements.

The cost breakdown also matters for agent deployments. The source says about 89% of the $0.25 task cost is input-related, including cache writes, while output costs less than three cents. A service that repeatedly sends large documents or code repositories may therefore spend more on input than the headline output-token price suggests. Actual bills will depend on prompt size, caching and provider implementation.

How the Benchmarks Compare

The release follows Nemotron 3 Super, which scored 13 on the same index, according to the supplied report. The new model’s reported cost per index task falls from $1.64 to $0.25. The report attributes the change partly to lower output volume per task: about 46,000 output tokens, compared with 114,000 for its predecessor. These are benchmark-run figures, not a promise of equivalent savings for every application.

The model’s results are uneven across evaluations cited in the source. Its GDPval-AA score, covering economically oriented tasks, was 1,162 Elo, compared with 490 for Nemotron 3 Super. AutomationBench-AA was 14.8%, up from 3.0% for Nemotron 3 Ultra, while the source describes that absolute rate as low. The reported AA-Omniscience hallucination rate fell from 87% for the predecessor to 36% for Super VL; that remains a benchmark-specific result and does not mean hallucinations are eliminated.

The licensing comparison also needs care. The source says Super VL leads among permissively licensed U.S. open-weight models, while same-day release Muse Spark Open scored 39 under a license described as commercially usable with restrictions. The distinction is relevant to organizations whose legal teams require broad commercial permissions; it does not make one license suitable for every deployment.

“Super VL is a very good engine and an average brain.”

— Thorsten Meyer, ThorstenMeyerAI.com

Provider Results Remain Variable

The reported 630-token-per-second result came from one pre-release endpoint, and the source does not establish that other providers or self-hosted setups will reproduce it. Across providers, the reported fifth-to-95th-percentile range was 106 to 534 tokens per second. Hardware, quantization, prompt length and service configuration may affect results, but the supplied material does not provide a controlled comparison isolating those factors.

It is also unclear how closely benchmark costs and scores will match typical production use. The source gives a per-task estimate and evaluation-specific results, but not independently verified costs for a range of customer workloads. Nor does the material establish how the model performs across different image types, languages or safety-sensitive applications. The reported hallucination reduction is substantial relative to the predecessor, but a 36% benchmark rate remains a limitation, not a guarantee of factual reliability.

Test Before Production Deployment

The next practical step for prospective users is to test their own prompts and workloads on the provider or hardware they intend to use. That should include measuring first-token delay, full response time, input and cache costs, accuracy on relevant tasks, and failure rates. Teams considering self-hosting will also need to evaluate hardware needs and operational costs; the supplied material does not quantify those requirements.

Artificial Analysis’ reported figures provide an initial comparison point, but they do not settle whether Super VL is the right choice for a particular product. The model’s case will depend on whether low latency, open weights and its license outweigh competitors’ higher benchmark scores or lower estimated task costs. Further provider measurements and independent production evaluations would help clarify that trade-off.

Key Questions

What is Nemotron 3.5 Super VL?

It is NVIDIA’s 120-billion-parameter open-weight vision-language model, with 12 billion active parameters in its mixture-of-experts design. It accepts text and images and produces text, according to the supplied report.

Does it deliver sub-second responses?

Artificial Analysis reported a 0.89-second median time to first token across six providers. The source separately says reasoning mode took about 11 seconds to produce its first answer token, so the sub-second figure should not be read as a guarantee of a complete response within a second.

Is it the best-value open-weight model?

Not on the comparisons in the supplied report. It lists other models with higher Intelligence Index scores for similar or lower estimated task costs. Super VL’s potential advantages include speed, its stated commercial-use license and self-hosting options.

Can businesses use the model commercially?

The supplied material says its OpenMDW-1.1 license allows commercial use. Businesses should review the license terms and their own deployment requirements before relying on that description.

What should teams verify before adopting it?

Teams should measure latency, task accuracy and total cost on their chosen provider or hardware, using their own prompt lengths and caching patterns. The published results vary across providers and may not match a specific production workload.

Source: ThorstenMeyerAI.com