The desk
Die Brief

DIE.BRIEF
Hardware desk
Models SINGLE-SOURCE

Nvidia's open Nemotron agent model moves to enterprise and edge

The Die Brief Desk, after github.com

Open weights are the lead, with enterprise agents first and two Bolt memory tiers behind it.

In the overall Artificial Analysis ranking, Super VL trails the leading closed models by a wide margin but beats Nvidia's own Nemotron 3 Ultra. | Image: Nvidi — Picture from the-de
In the overall Artificial Analysis ranking, Super VL trails the leading closed models by a wide margin but beats Nvidia's own Nemotron 3 Ultra. | Image: Nvidi — Picture from the-de
Nemotron 3.5 Super VLNemotron 3.5 Bolt 74B-A9B 24GBNemotron 3.5 Bolt E4B 4GB
parameters120 billion74B-A9BE4B
memory24GB4GB
ship dateavailable nowavailable nowOct. 23
benchmark5x faster
size comparisonroughly one-fifth of largest reasoning model

The 120-billion-parameter open model is built for continuous enterprise agents, and only 12 billion parameters are active per token.

The enterprise buyer gets open weights, but the page ties the models to Nvidia infrastructure and local Nvidia systems.

Super VL is slightly less accurate than competitors on the coding benchmark, but Nvidia says it finishes the tasks four times faster. | Image: Nvidia — Picture from the-decoder.com
Super VL is slightly less accurate than competitors on the coding benchmark, but Nvidia says it finishes the tasks four times faster. | Image: Nvidia — Picture from the-decoder.com

The 24GB Bolt model is available now, and the 4GB version is scheduled for Oct. 23.

At a comparable Intelligence Index score, Super VL generates up to five times more tokens per second than other open models in its class, according to Nvidia. | Image: Nvidia — Pic
At a comparable Intelligence Index score, Super VL generates up to five times more tokens per second than other open models in its class, according to Nvidia. | Image: Nvidia — Pic

The accuracy claim is a one-fifth-size comparison with Nvidia's largest reasoning model, not a proof that Super VL beats every closed frontier model.

On the DGX Spark, no other tested model was both more accurate and faster than Super VL. | Image: Nvidia — Picture from the-decoder.com
On the DGX Spark, no other tested model was both more accurate and faster than Super VL. | Image: Nvidia — Picture from the-decoder.com

The 5x speed figure, independent accuracy, and throughput outside Nvidia's own claims remain unmeasured.

The token cost per agent task, latency on local systems, and edge-model accuracy also stay open.

The useful shift is open enterprise agents, but the Nvidia stack still shapes the cost.

After github.com. We did not report this. The pictures, if any, are theirs.

October 15, 2026 · 1 min
Copy link Share to X

Also

Apple benchmarks Muse Spark Open on four Mac Studios via MLX distributed inference

CONFIRMED · October 15, 2026

Google Antigravity screens UniProt for antimicrobial peptide candidates

CONFIRMED · October 15, 2026

NVIDIA splits Nemotron agents by memory tier

CONFIRMED · October 15, 2026