NVIDIA Nemotron 3.5: what the local tests actually show

Two downloads, one more on the calendar

NVIDIA’s 15 October announcement makes Super VL and Bolt 74B available now. The smaller Bolt E4B, targeting 4GB, is scheduled for 23 October. That last model is still a date in the diary.

The Super VL model card describes a vision-and-language model with 120 billion total parameters and 12 billion active per token. It accepts text, images and video and produces text. Bolt is compressed from that family to 74 billion total parameters, nine billion active, with an approximately two-bit quantized release. Its GGUF card also supports images; it is not simply the text-only sibling.

The 24GB label describes Bolt’s deployment target. It does not promise that a 24GB computer can run the model, a long conversation and every other application comfortably. Model weights, context storage and the software serving requests all have to fit. Super VL advertises up to a million tokens of context, but its deployment recipes vary with available memory and concurrency.

NVIDIA introduces the laptop-oriented Bolt below. The source post establishes the release; it does not independently establish how quickly your workload will run.

Compression changes the capability budget

Artificial Analysis’s 15 October evaluation puts Super VL at 26.5 on its Intelligence Index and Bolt at 21.4. The smaller package gives up measurable capability on that composite evaluation. Those scores are neither percentages of tasks completed nor a forecast for a particular company’s documents.

The same evaluation reports roughly 46,000 output tokens per task for Super VL and 82,000 for Bolt. A smaller model can do more work before reaching an answer. Parameter count alone therefore cannot tell a team how long a job will take or what a hosted call will cost.

The comparison below is deliberately dated. It compares these two releases on one index; it is not a current leaderboard or a declaration that either model leads every task.

The laptop result has a laptop attached

Corey Noles’s firsthand review for The Neuron tested Bolt’s IQ2_XXS GGUF through Unsloth Studio on a Surface Laptop Ultra with RTX Spark and 128GB unified memory. Super VL ran through OpenRouter Chat. Those are different deployment environments, not a controlled race between two models on one machine.

Bolt reached 65.5 tokens per second on a three-box logic puzzle. More revealing was the writing exercise: a flexible-length request took 47 seconds and 2,029 generated tokens; a related request with exact word counts took 137 seconds and 6,797 tokens. The requirements changed, so this does not isolate a single cause. It does show why a headline speed figure can miss the time spent satisfying instructions.

Noles’s practical observation is useful for drafting workflows, where a rigid constraint may cost more effort than the first version of the brief suggests.

A fresh Spark test exposes the settings

A separate test published on 17 October by engineer netrunner reports Super VL moving from about 75.4 to 91.0 tokens per second on one DGX Spark. The linked TensorFold recipe supplies patches, benchmark settings and reply checks. It is the recipe author’s own measurement, with inspectable material rather than independent replication.

The headline run generates a predictable count sequence with greedy decoding, thinking disabled and a 400-token limit. All 24 measured replies matched the stock output. But the recipe and stock headline runs use different context limits, so the whole gain cannot be assigned to the patches alone. The repository also includes separately matched-setting tests and concurrency measurements.

This is encouraging engineering evidence for that machine and configuration. It does not make a DGX Spark result a laptop result, or a counting workload a forecast for document review. The original post below is worth reading together with its linked settings.

Control starts after the download

The released models use OpenMDW 1.1, which permits commercial use subject to its terms, including notices and the agreement when redistributing covered material. Downloadable weights give a team deployment choices; they do not configure data retention, access permissions or where its serving infrastructure runs.

For a European organisation considering local document or creative work, those are concrete decisions: choose the hardware and runtime, test the actual files and instructions, then measure completed jobs rather than tokens alone. A hosted trial can help explore a model, but it does not establish the behaviour of a later local installation.

Nemotron’s useful advance here is a wider set of deployable sizes. The evidence already shows why the smaller box and the larger box deserve separate trials. The next question is how each handles repeated real tasks, long contexts and simultaneous users under the settings a team intends to keep.