Skip to content
archdiff

models / compare / gemma-4-vs-nemotron-3-ultra

Gemma 4 31B vs Nemotron 3 Ultra

How Gemma 4 31B (Google DeepMind, April 2026) and Nemotron 3 Ultra (NVIDIA, June 2026) are built, drawn on the same November 2026 blueprint. They make different choices on 19 of 29 components. Select a component in either diagram to find it in the table.

--- frontier-llm@2026-11+++ gemma-4@2026-04+2~7−6

Gemma 4 31B

Google DeepMind, April 2026, 30.7B

Text · code · tool outputText · code · tool outputImages · video · audioImages · video · audioSentencePiece BPE with byte fallback — Input & outputSentencePiece BPE with byte fallbackInput & outputVision / audio encoder — native ViT or encoder-free patchesVision / audio encodernative ViT or encoder-free patchesToken embedding — media embeds spliced in placeToken embeddingmedia embeds spliced in placeEngram n-gram memory — hashed 2–4-gram lookup (V4.1)Engram n-gram memoryhashed 2–4-gram lookup (V4.1)Single residual stream with per-layer scalars — Residual & normSingle residual stream with per-layer…Residual & normRMSNorm — pre-normRMSNormpre-normQK-Norm — per-head RMSNorm on q, kQK-Normper-head RMSNorm on q, kRoPE by layer type: full RoPE local, p-RoPE global — AttentionRoPE by layer type: full RoPE local, p…AttentionGated DeltaNet / KDA — linear · ~3 of every 4 layersGated DeltaNet / KDAlinear · ~3 of every 4 layersGlobal GQA with keys reused as values — AttentionGlobal GQA with keys reused…AttentionSparse / compressed attn — DSA indexer · CSA / HCASparse / compressed attnDSA indexer · CSA / HCASliding-window attention — local · 128–1,024-token windowSliding-window attentionlocal · 128–1,024-token windowOutput gate + sink — σ(x·W_g) ⊙ o · learnable sinkOutput gate + sinkσ(x·W_g) ⊙ o · learnable sinkPost-attention RMSNorm (sandwich norm) — Residual & normPost-attention RMSNorm (sandwich norm)Residual & norm⊕ residual⊕residualRMSNorm — pre-normRMSNormpre-normDense GeGLU feed-forward — Mixture of expertsDense GeGLU feed-forwardMixture of expertsPost-feed-forward RMSNorm (sandwich norm) — Residual & normPost-feed-forward RMSNorm (sandwich no…Residual & normRouted experts — top-6–22 of 256–896 · SwiGLURouted expertstop-6–22 of 256–896 · SwiGLUShared expert — 1–2 · always activeShared expert1–2 · always active⊕ residual⊕residualFinal RMSNormFinal RMSNormTied LM head with logit soft-capping — Input & outputTied LM head with logit soft-cappingInput & outputAutoregressive MTP drafter (4 layers, reads trunk KV) — Input & outputAutoregressive MTP drafter (4 layers,…Input & outputNext-token logitsNext-token logitsDraft tokensDraft tokensMuon optimizer — Newton–Schulz · QK-Clip · per-headMuon optimizerNewton–Schulz · QK-Clip · per-headFP8 → FP4 training — block-scaled FP8 · MXFP4/NVFP4 QATFP8 → FP4 trainingblock-scaled FP8 · MXFP4/NVFP4 QATLong-context extension — staged 4K → 1M · YaRNLong-context extensionstaged 4K → 1M · YaRNRL with verifiable rewards — GRPO family · async agentic RLRL with verifiable rewardsGRPO family · async agentic RLOn-policy distillation — reverse KL to specialist teachersOn-policy distillationreverse KL to specialist teachersSpeculative decoding — MTP self-drafting · DSparkSpeculative decodingMTP self-drafting · DSparkKV-cache compression — FP4 KV · cross-layer reuse · SSDKV-cache compressionFP4 KV · cross-layer reuse · SSD
AttentionMixture of expertsResidual & normInput & outputTraining & inferenceData flowResidual / skipAddedReplacedNot usedNot disclosed

--- frontier-llm@2026-11+++ nemotron-3-ultra@2026-06+0~4−7

Nemotron 3 Ultra

NVIDIA, June 2026, 550B total, 55B active

Text · code · tool outputText · code · tool outputImages · video · audioImages · video · audioByte-level BPE tokenizer — 128K–262K vocabByte-level BPE tokenizer128K–262K vocabVision / audio encoder — native ViT or encoder-free patchesVision / audio encodernative ViT or encoder-free patchesToken embedding — media embeds spliced in placeToken embeddingmedia embeds spliced in placeEngram n-gram memory — hashed 2–4-gram lookup (V4.1)Engram n-gram memoryhashed 2–4-gram lookup (V4.1)Single residual stream, one sublayer per layer — Residual & normSingle residual stream, one sublayer p…Residual & normRMSNorm — pre-normRMSNormpre-normQK-Norm — per-head RMSNorm on q, kQK-Normper-head RMSNorm on q, kPartial RoPE / NoPE — rotate 64 dims, or nonePartial RoPE / NoPErotate 64 dims, or noneMamba-2 selective state-space layers — AttentionMamba-2 selective state-spac…AttentionGlobal GQA attention without positional encoding — AttentionGlobal GQA attention without…AttentionSparse / compressed attn — DSA indexer · CSA / HCASparse / compressed attnDSA indexer · CSA / HCASliding-window attention — local · 128–1,024-token windowSliding-window attentionlocal · 128–1,024-token windowOutput gate + sink — σ(x·W_g) ⊙ o · learnable sinkOutput gate + sinkσ(x·W_g) ⊙ o · learnable sink⊕ residual⊕residualRMSNorm — pre-normRMSNormpre-normMoE router — σ / √softplus · bias balancingMoE routerσ / √softplus · bias balancingLatentMoE routed experts (ReLU², 2,048-dim latent) — Mixture of expertsLatentMoE routed experts (ReLU², 2,048…Mixture of expertsShared expert — 1–2 · always activeShared expert1–2 · always active⊕ residual⊕residualFinal RMSNormFinal RMSNormLM head — untied · vocab logitsLM headuntied · vocab logitsMTP module(s) — predict t+2 … t+kMTP module(s)predict t+2 … t+kNext-token logitsNext-token logitsDraft tokensDraft tokensMuon optimizer — Newton–Schulz · QK-Clip · per-headMuon optimizerNewton–Schulz · QK-Clip · per-headFP8 → FP4 training — block-scaled FP8 · MXFP4/NVFP4 QATFP8 → FP4 trainingblock-scaled FP8 · MXFP4/NVFP4 QATLong-context extension — staged 4K → 1M · YaRNLong-context extensionstaged 4K → 1M · YaRNRL with verifiable rewards — GRPO family · async agentic RLRL with verifiable rewardsGRPO family · async agentic RLOn-policy distillation — reverse KL to specialist teachersOn-policy distillationreverse KL to specialist teachersSpeculative decoding — MTP self-drafting · DSparkSpeculative decodingMTP self-drafting · DSparkKV-cache compression — FP4 KV · cross-layer reuse · SSDKV-cache compressionFP4 KV · cross-layer reuse · SSD
AttentionMixture of expertsResidual & normInput & outputTraining & inferenceData flowResidual / skipAddedReplacedNot used

Component by component

Gemma 4 31B and Nemotron 3 Ultra differ on 19 of 29 components.

Blueprint componentGemma 4 31BNemotron 3 Ultra

SentencePiece BPE with byte fallbackGemini's 262,144-entry SentencePiece vocabulary: character-level BPE, split digits, bytes only as a fallback.

Same as the blueprintByte-level BPE with 131,072 entries, the first 1,000 of which are reserved special and control tokens; the pre-tokenizer splits every digit into its own piece. The model is text-only.

Same as the blueprintA ~550M-parameter ViT (27 layers, width 1,152, MLP 4,304, 16 heads, 16-px patches) with 2D RoPE and 2D absolute position embeddings, kept frozen during pretraining. Its outputs are pooled 3 × 3, so each soft token covers 48 × 48 px, under a per-image budget of 70, 140, 280 (default), 560 or 1,120 tokens. Video enters as frames; the 31B has no audio encoder.

Not used

Single residual stream with per-layer scalarsOne d = 5,376 stream, RMSNorm before and after each sublayer, and a stored scalar that rescales the stream after every block.

Single residual stream, one sublayer per layer108 pre-norm layers on one d = 8,192 residual stream; each layer is one mixer or one MoE, not an attention+FFN pair.

Same as the blueprintPer-head RMSNorm with a learned gain on queries and keys (256-dim heads in local layers, 512-dim in global ones), so the logits use scale 1 instead of 1/√d. Values get a gain-free RMSNorm. There is no attention-logit soft-cap.

Not used

RoPE by layer type: full RoPE local, p-RoPE globalLocal layers rotate all 256 dims (θ = 10⁴); global layers keep only the 64 fastest of 256 frequency pairs (θ = 10⁶).

Same as the blueprintNoPE: the 12 attention layers apply no positional encoding (Nemotron 3 Super omits positional embeddings, and Ultra keeps the same architecture). Token order reaches them only through the Mamba-2 layers' causal convolutions and recurrent state. config.json still lists rope_theta = 10,000, but the Hugging Face implementation never applies it.

Not used

Mamba-2 selective state-space layers48 of the 60 mixer layers: SSD with 256 heads × 64 dims, state size 128, 8 B/C groups, causal conv of kernel 4.

Global GQA with keys reused as valuesEvery sixth layer: full causal attention, 32 query heads and 4 KV heads of dim 512, V taken from the key projection.

Global GQA attention without positional encoding12 layers of dense causal attention: 64 query heads share 2 KV heads (d_h = 128); no RoPE, QK-norm, gate or sink.

Same as the blueprint50 of 60 layers, in runs of five between global layers. Each uses a 1,024-token window with 32 query heads and 16 KV heads of dim 256 (2 query heads per KV head), separate K and V projections and full RoPE with θ = 10,000. Their KV cache is bounded at 1,024 tokens per layer.

Not used

Post-attention RMSNorm (sandwich norm)The attention output is RMS-normalized with a learned gain before the residual add; there is no output gate or sink.

Not present

Dense GeGLU feed-forwardOne gated MLP per layer, 5,376 → 21,504 → 5,376 with tanh-approximated GELU; all 30.7B parameters are active.

Same as the blueprintLinear router at the full 8,192 width: sigmoid scores over 512 experts, a per-expert bias steers top-22 selection, and the selected scores are renormalized and scaled by 5.0. Balancing follows the Nemotron 3 Super recipe: aux-loss-free bias updates (rate 1e-3) plus a standard load-balancing loss (coefficient 1e-4).

Post-feed-forward RMSNorm (sandwich norm)The MLP output is RMS-normalized with a learned gain before it is added to the residual stream.

Not present

Not used

LatentMoE routed experts (ReLU², 2,048-dim latent)512 non-gated squared-ReLU experts of width 5,120 run in a 2,048-dim latent space; 22 active per token.

Not used

Same as the blueprintOne shared expert per MoE layer at full width: a non-gated squared-ReLU MLP, 8,192 → 10,240 → 8,192, added to the routed output without a router weight. The released checkpoint stores it in FP8 (per-tensor), not NVFP4.

Tied LM head with logit soft-cappingLogits come from the input embedding matrix (no separate head) and are squashed by 30·tanh(z/30).

Same as the blueprintUntied 8,192 × 131,072 output projection, shared with the MTP block and kept in BF16. Its gradient accumulation must stay in FP32: lowering it to BF16 erased the MTP heads' small gradient contribution and caused a pretraining divergence at about 8T tokens.

Autoregressive MTP drafter (4 layers, reads trunk KV)A separate ~500M-parameter, 4-layer Transformer drafts tokens one at a time while attending to the 31B's own KV cache.

Same as the blueprintOne MTP block (an attention layer followed by a LatentMoE layer) whose weights are shared across steps and applied recursively. It ran 2 steps in pretraining (total loss weight 0.1, 0.05 per step). MTP Boosting then trains only this head with a forward-KL loss to the frozen backbone's logits over 7 steps, to match the inputs it sees when drafting.

Not disclosed

Not used

Not disclosed

Same as the blueprintPretraining used 8,192-token sequences (Nemotron 3 Super recipe). A 33B-token continued-pretraining phase then trained at 1,048,576 tokens for 92% of iterations and at 4,096 for the rest, at a constant LR of 2.5e-6. With no RoPE, no YaRN-style rescaling is involved.

Not disclosed

Same as the blueprintAfter a two-stage SFT, a single unified RLVR stage ran asynchronous GRPO across agentic, coding, math, STEM, safety, chat, instruction-following and long-context environments. The maximum generation length grew from 48K to 64K tokens, and RLHF used a released GenRM. MTP drafting with 5 tokens sped up rollout generation 1.46×.

Not disclosed

Same as the blueprintMulti-teacher on-policy distillation (MOPD) in two iterations. More than ten domain-specialized teachers (STEM, coding, SWE, search, terminal use, office work, chat, safety and others) give dense token-level reverse-KL signals on the student's own rollouts. The RLVR student acts as a self-teacher for uncovered domains.

More comparisons