Home›Models›nemotron-3.5-super-120b
nemotron-3.5-super-120b
Architecture mixture of experts88 LayersLicense openmdw-1.1Maximum context 1,024kVocabulary 131,072
At Q4_K_M it is a 65.29 GiB file, and it fits on 5 of the 18 machines we measure (the one with the least memory: Xiaomi AI Cube, 80 GB — engineering prototype). Here the file is what fills the memory: even at its maximum context of 1024k tokens, the context adds 8.00 GiB — and the same content in German takes ×1.40 the tokens of English.
Model size at Q4_K_M
Measured by Local AI Scope
Context memory at 8k tokens
Measured by Local AI Scope
What German costs versus English
Measured by Local AI Scope
Videos: nemotron-3.5-super-120b 2
Official footage and the best walk-throughs we found. Press play and it opens right here.
The videos live on YouTube, and nothing loads from there until you press play (no-cookie mode). The thumbnail is served by us.
Sizes of the real published files
These are the sizes of the files actually published by the author, added up when the model ships split into parts. Not an estimate from the parameter count.
| Compression | Model size |
|---|---|
Q4_K_M |
65.29 GiB |
This model also reads images. For that it needs a second file, the vision projector (mmproj-NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16.gguf, 1.61 GiB), published next to the model and not included in the figures on this page: they count text only. If you are going to send it images, add it to the model size.
Context memory
Computed from the attention pattern this model declares layer by layer, not from the generic formula. That is why the figure is often far smaller than other calculators tell you.
- 8k tokens → 0.06 GiB
- 32k tokens → 0.25 GiB
- 128k tokens → 1.00 GiB
What it costs in each language
Measured by running this model’s own tokenizer over the same content in four languages. Tokens are what you pay for in context, in memory and in your cloud bill.
| Kind of text | EN | ES | FR | DE |
|---|---|---|---|---|
| Code | — | ×1.23 | ×1.31 | ×1.51 |
| Contracts | — | ×1.21 | ×1.25 | ×1.37 |
| Business email | — | ×1.19 | ×1.22 | ×1.31 |
| Support tickets | — | ×1.28 | ×1.36 | ×1.37 |
English is the baseline: every figure is how many times more tokens the same content costs in that language. Average across the whole corpus: DE ×1.40.
Where it fits
Fits means the model, its context memory and the working headroom all fit at the stated context, leaving the system its share. 5 of the 18 machine profiles tested (Q4_K_M, 8k tokens).
Before you hit download: as of 17 October 2026, this model is so new that only the latest versions open it. llama.cpp opens it from its stable v0.7.0 (build b11669, released on 15 October 2026); earlier versions do not. Heads-up: since that version its server listens on port 9931 by default, not 8080 (the Docker images keep 8080). Ollama has offered it since 0.40.5 (ollama run nemotron-3.5-super), but on a PC its q4_K_M is not the file in the table: it weighs 82.35 GiB, plus the 1.61 GiB vision projector that Ollama loads every time, 83.96 GiB in all. Same name, different recipe: in this model part of the expert tensors cannot use the Q4_K format, and each file falls back to a different one (IQ4_NL in LM Studio’s, Q5_0 and Q8_0 in Ollama’s), as we read in each file’s header. With Ollama’s file, the Xiaomi AI Cube (80 GB) can no longer hold the model and the PC with Ryzen AI Max+ 395 drops from 1024k to 256k of context; with LM Studio’s (65.29 GiB), the Xiaomi does hold it, up to 256k, as the table says. LM Studio offers that GGUF from its community program (lmstudio-community), published on 14 October 2026; we have not checked which version of its engine opens it, so update it before you try. On a Mac, Ollama’s catalogue lists an MLX version under that same name (nvfp4, 74.09 GiB); mlx-lm on its own we have not been able to confirm. Updated on 19 October 2026: NVIDIA’s vLLM recipe calls for 0.32.0, stable since 18 October 2026, or a nightly build. It reads images through the 1.61 GiB projector, which the table does not add; NVIDIA says it also understands video (clips of up to 2 minutes), but in its vLLM and SGLang recipes: Ollama and llama.cpp announce images only, for now. There is also a Q8_0 (119.65 GiB), but our catalogue does not list it, so the Model Finder will not suggest it: by our count, text only, it would fit on the PC with Ryzen AI Max+ PRO 495 and the Mac Studio with M5 Ultra up to 1024k, and on the Mac Studio with M5 Max and 128 GB up to 128k.