NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base

2026-10-15

Date the weights (or, if earlier, the technical report) were first made public, YYYY-MM-DD, as dated by the publisher. A Hugging Face repository creation date is used only when no dated publisher page exists, with a note.

other (openmdw-1.1)

License of the released weights as declared by the publisher (model card metadata or license file), e.g. apache-2.0, mit, or the name of a custom license.

text + image + video

Input modalities of the released checkpoint as the publisher states them, from: text, image, video, audio, joined with ' + ' in that order ("text", "text + image + video").

120B parameters

Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.

12B parameters

Parameters used per token in the forward pass, as reported by the publisher.

Not disclosed

Active parameters excluding the input embedding and the output head (unembedding). Stated by the publisher, or computed from params.active by subtracting vocab × d_model for each of the embedding and output head that params.active is stated to include (basis inferred, note with the arithmetic). Not disclosed when the scope of params.active is not stated.

120.9B parameters

Calculated: embedding + output head + all experts + other weights, nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16@5d80455 safetensors headers; reproduces stated total (backbone) and active (non-embedding)

Parameters of the text backbone counted from the released checkpoint's safetensors headers (embedding, output head, all experts and other weights; excluding MTP modules, vision/audio towers and n-gram lookup tables). Shown only when the counts reproduce the publisher's stated total and active parameters within 5% under some scope.

1.074B parameters

Calculated: input embedding + output head, nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16@5d80455 safetensors headers

Input embedding plus output head (once if tied), counted from the checkpoint.

119.8B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16@5d80455 safetensors headers; reproduces stated total (backbone) and active (non-embedding)

Language backbone weights excluding input embeddings, output head, lookup/Engram tables, MTP modules and vision/audio towers. Equals params.counted_total minus params.counted_embedding. Same validation as params.counted_total.

2.873B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16@5d80455 safetensors headers

Largest group of language-backbone weights sharing a full layers.<index> tensor path, including every routed expert. Excludes lookup/Engram tables, embeddings, output head, MTP and other modalities. Hybrid checkpoints may store mixers and FFNs as separate layers. All-gather volume also depends on wrapping, sharding and dtype.

11.89B parameters

Calculated: non-expert weights + routed experts × 22/512, excluding embedding and output head, nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16@5d80455 safetensors headers; reproduces stated total (backbone) and active (non-embedding)

Active parameters per token excluding the input embedding and output head, counted from the checkpoint as all non-expert weights plus routed-expert weights × top_k / experts. Same validation as params.counted_total.

Rounded publisher figures for the multimodal model; whether the 120B includes the C-RADIO encoder, connector, MTP head and embeddings is not stated (§4.2 lists all three under the 120B model). The checkpoint index totals 248,906,833,944 bytes (about 124.5B parameters at 2 bytes each), including the vision tower, connector (mlp1) and MTP block; the predecessor backbone alone is 120.6B.

What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).

48

Inferred from the source; see the explanation below.

Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.

4,096

Residual stream width.

131,072

Embedding table rows (vocab_size in config, may include padding).

No

Input embedding and output head share weights.

85.3333

Calculated: d_model / layers

Width-to-depth ratio, d_model divided by the number of blocks.

GQA, Mamba-2

Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).

40 Mamba-2 + 8 GQA

Inferred from the source; see the explanation below.

Token-mixer order, summarized as a pattern or layer positions. The fact's optional sequence lists each mixer in order; a ratio alone does not establish the order. Auxiliary prediction modules and standalone FFNs are excluded.

8

Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.

5

Calculated: attn.linear_layers / attn.global_layers

attn.linear_layers divided by attn.global_layers (3 for '3 Gated DeltaNet : 1 GQA').

40

Layers with a fixed-size recurrent state (linear attention, DeltaNet variants, Mamba).

32

Query heads in the global-attention layers.

2

Key/value heads in the global-attention layers. Not applicable to MLA, whose cache is a shared latent.

4,096

Calculated: heads × head_dim

attn.heads × attn.head_dim, the query projection width of global layers.

16

Calculated: heads / kv_heads

attn.heads divided by attn.kv_heads in global layers.

128

Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).

NoPE

Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.

No

Queries and keys are normalized per head before the dot product.

No

Query/key/value projections have bias terms.

Mamba-2: 128 heads × head dim 64, state 128, 8 groups, conv kernel 4

Configuration of the linear / SSM layers (e.g. Gated DeltaNet heads and dims, Mamba-2 state size), as one string.

128

Value heads (or SSM heads) in the linear / SSM layers.

64

Head dimension of the linear / SSM layers (value head dim for DeltaNet variants, head_dim for Mamba-2).

4

Kernel size of the causal short convolution in linear / SSM layers or attention projections.

ReLU²

FFN activation (SwiGLU, GeGLU, ...).

No

FFN projections have bias terms.

512

Routed experts per MoE layer.

22

Routed experts selected per token.

1

Always-on experts per MoE layer.

2,688

Intermediate width of one routed expert.

5,376

Total intermediate width of the shared expert(s).

0

Number of initial blocks that use a dense FFN instead of MoE.

sigmoid

Gating function over expert logits (softmax, sigmoid).

sequence-level load-balancing loss 1e-5 during distillation; expert-selection bias (e_score_correction_bias) in the router

Load-balancing method, with coefficients where disclosed.

5

Multiplier applied to the combined routed-expert output (routed_scaling_factor).

Yes

Selected expert weights are renormalized to sum to 1.

1,024

Width of the projected space routed experts operate in (LatentMoE), when smaller than d_model.

4

Calculated: d_model / moe.latent_dim

shape.d_model divided by moe.latent_dim, how much LatentMoE compresses the hidden state before routed experts.

0.65625

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

23.2727

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.1

Calculated: params.active / params.total

params.active divided by params.total.

2

Inferred from the source; see the explanation below.

Additional future tokens predicted during pretraining (0 when the model was trained without MTP).

1

MTP modules with distinct weights. 1 with mtp.depth > 1 means one module is applied recursively.

next-token embedding and hidden state each normed (enorm, hnorm), concatenated and projected 2×4096→4096 (eh_proj); then one GQA attention layer and one LatentMoE layer; final norm

Contents of one MTP module (mixer and FFN type, norms, how the previous hidden state and next-token embedding are combined).

MoE

FFN in the MTP module: 'dense' or 'MoE', with its width when it differs from the backbone.

Yes

Inferred from the source; see the explanation below.

MTP modules use the main model's input embedding and output head.

0.1

Inferred from the source; see the explanation below.

Weight of the MTP loss, including its schedule.

Yes

The released checkpoint contains the MTP module weights.

RMSNorm

Normalization layer type in the blocks.

pre

Pre-norm, post-norm, or sandwich (pre and post).

1e-05

Epsilon of the block normalization layers.

Not disclosed

Sequence length of the main pretraining stage.

1.049M tokens

Longest sequence length trained on after context extension.

1.049M tokens

Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.

1M-context CPT at constant LR 4.5e-6 for 35B tokens (Nemotron 3 Ultra long-context blend) after Indic CPT; 65.4B-token mid-training at 256K dropped RULER-1M from 79.9 to 33.5, so a 9B-token 1M recovery stage with the same recipe followed (RULER-1M 81.0)

How context was extended (stages, RoPE base change, YaRN factor).

1.209T tokens

Inferred from the source; see the explanation below.

Tokens seen in pretraining including annealing and context extension, as reported.

Not disclosed

Optimizer used for the bulk of parameters.

4.5e-05

Peak learning rate of the main pretraining stage.

4.5e-06

Learning rate at the end of the decay.

Distillation: linear ramp 4.5e-6 → 4.5e-5 over 200B tokens, linear decay back to 4.5e-6 over the remaining 800B; long-context CPT constant 4.5e-6; mid-training reset, linear warmup to 1e-5 then constant; Indic CPT LR not stated

Inferred from the source; see the explanation below.

Learning-rate schedule shape with phase boundaries.

200B tokens

Inferred from the source; see the explanation below.

Warmup length in steps or tokens.

Not disclosed

Global batch size, including any batch-size ramp: sequences × sequence length (token total), or the token total alone when only that is stated.

Not disclosed

Decoupled weight decay coefficient.

Not disclosed

(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.

Not disclosed

Global gradient-norm clipping threshold.

Not disclosed

Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.

Backbone from Nemotron 3 Super Base; MTP head from the first of Nemotron 3 Super Base's two MTP heads (pre-long-context checkpoint), second discarded; C-RADIO vision encoder and connector added after text pretraining

Weight initialization (std or scheme) as stated by the publisher; config initializer_range alone is not evidence.

Logit distillation from Nemotron 3 Ultra Base: top-128 teacher logits plus a sink token for residual mass, Skew Reverse KL (α = 0.1) combined with cross-entropy (0.9 / 0.1)

Objectives beyond next-token prediction and MTP (fill-in-the-middle rate, distillation, multimodal tokens from the start).

Distillation 1T tokens (recipe of Lightning 3.5, which used Nemotron 3 Ultra phase-2 data); Indic CPT 100B tokens: 10% synthetic (Gemma-4-31B-it translations of Nemotron-CC) + natural Indic text, 90% preceding mixture; long-context CPT on the Nemotron 3 Ultra LC blend; agentic mid-training: issue–PR 41%, repositories 20%, agentic traces 10%, commit diffs 3%, build/CI logs 1%, phase-2 pretraining replay 20%, long-context data 5%

Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).

December 2025

Pretraining data cutoff date.

GB300

Accelerators used for pretraining, count and duration where stated.

GB300

Accelerator count and type for the main pretraining run, as the count with thousands separators and the chip's short name ("2,048 H800", "2,048 GH200"). A stated range or bound stays as stated ("2,048–4,096 GH200").

TP8 · EP64 · CP16

Parallelism layout of the main pretraining run as stated, as degree-annotated terms in the order TP, PP, VPP, EP, CP, DP ("PP16 · EP64 · ZeRO-1 DP"); a term without a stated degree has none ("FSDP").

• mtp-warmup

  Tokens

  100000000000

  Source: paper · PDF pages 4-10; §2.1-2.5

  we conduct a dedicated MTP warmup phase for 100 B tokens, freezing the backbone parameters and optimizing only the MTP head.

  MTP head only, backbone frozen; hidden-state mixing across recursive steps.

• distillation

  Tokens

  1000000000000

  Lr

  4.5e-05

  Source: paper · PDF pages 4-10; §2.1-2.5

  then distilled each model for 1 trillion tokens.

  Teacher Nemotron 3 Ultra Base; LR ramps 4.5e-6 → 4.5e-5 over 200B, decays linearly to 4.5e-6; lr is the peak.

• indic-cpt

  Tokens

  100000000000

  Source: paper · PDF pages 4-10; §2.1-2.5

  Super’s Indic continued-pretraining run has a nominal 100B-token budget

  10% Indic (synthetic + natural), 90% preceding mixture.

• long-context

  Tokens

  35000000000

  Seq len

  1048576

  Lr

  4.5e-06

  Source: paper · PDF pages 4-10; §2.1-2.5

  The LC-Phase was trained for 21B tokens for Lightning and 35B tokens for Super.

  Constant LR; CP16 · TP8 · EP64 on GB300.

• mid-training

  Tokens

  65400000000

  Seq len

  262144

  Lr

  1e-05

  Source: paper · PDF pages 4-10; §2.1-2.5

  We mid-trained the long-context checkpoint for 65.4B tokens at a sequence length of 256K.

  Agentic software-engineering blend; LR reset, linear warmup to 1e-5, then constant.

• long-context-recovery

  Tokens

  9000000000

  Seq len

  1048576

  Lr

  4.5e-06

  Source: paper · PDF pages 4-10; §2.1-2.5

  for which we trained 9B tokens with the same recipe

  Same recipe as the long-context phase; restores RULER-1M from 33.5 to 81.0.

Ordered pretraining stages. Each item: name, tokens, seq_len, lr, batch_tokens, batch_sequences, steps, note, quote, and source when it differs from the fact's source.

Distillation LRs anchored to Nemotron 3 Super Base: its final LR (4.5e-6) as minimum and 10% of its peak (4.5e-5) as maximum, to limit early drift from the initialization

How peak LR, batch size or other hyperparameters were chosen for the released run (fitted scaling law with its formula, μP transfer from a proxy width, sweep), as stated.

10.0783

Calculated: train.tokens / params.total

train.tokens divided by params.total.

101.756

Calculated: train.tokens / params.counted_active_non_embedding

train.tokens divided by params.counted_active_non_embedding, or by the stated params.active_non_embedding when there is no validated checkpoint count.

8,192 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 8 × (512 × 16/8) B/entry

Retained KV cache divided by context length, including bounded SWA caches and indexer keys. Exports use uniform BF16 at 131,072 tokens. The website can select uniform payload precision (no quantization overhead) or sourced component formats (including scales and documented padding). This is cache payload, not peak prefill memory or persistent prefix storage.

8,192 bytes

Calculated: 8192 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 8 × (512 × 16/8) B/entry

Long-context global KV bytes per token, including indexer keys; excludes bounded SWA caches. Compression rates and cross-layer cache reuse are included. Exports use uniform BF16; the website's storage selector also applies here. This is the field to compare with global-KV figures in reports, using the matching storage recipe.

683 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 8 × (512 × 16/8) B/entry

kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).

89,478,485 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 8 × (512 × 16/8) B/entry

kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).

1,073,741,824 bytes

Calculated: 8192 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 8 × (512 × 16/8) B/entry

Retained KV cache for one sequence at the selected context, including SWA windows and indexer keys. Exports use uniform BF16 at 131,072 tokens. Excludes recurrent state, compressor tail state, draft/MTP caches, temporary prefill tensors, allocator metadata and persistent prefix snapshots. Does not estimate peak prefill or total serving memory.

26

Artificial Analysis Intelligence Index v4.3.2 of the post-trained model, from a dated Wayback snapshot of its Artificial Analysis page. The note names the variant (reasoning setting).