Apertus-v1.5-70B-base
Swiss AI · base checkpoint
Open in the comparison table Hugging Face ↗ Paper ↗
KV calculations use one sequence at 131,072 tokens and 16-bit cache elements. They include bounded sliding-window caches. This is retained cache payload, not peak prefill memory. The comparison table also offers sourced mixed-precision layouts, including quantization scales.
Release
- Released
2026-07-24
Source: blog · EPFL AI Center news, Apertus 1.5: Building the next generation of open AI infrastructure
Published 24 July, 2026 EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) today released Apertus 1.5
Release of the Apertus 1.5 suite on Hugging Face; the separate -base repository was created 2026-09-23 and the report is dated 2026-10-11.
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.
- License
apache-2.0
Source: model-card · README.md; YAML front matter
license: apache-2.0
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.
- Input modalities
text + image + audio
Source: paper · p.6 §1
The models accept image and audio inputs in addition to text
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").
Parameters
- Total params
70B parameters
Source: model-card · README.md; intro
Apertus 1.5 is a family of 8B and 70B fully open, multilingual, and multimodal models
Size designation; no exact count is reported.
Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.
- Active params
70B parameters
Inferred from the source; see the explanation below.
Source: model-card · README.md; intro
Apertus 1.5 is a family of 8B and 70B fully open, multilingual, and multimodal models
Dense model; active equals the stated size.
Parameters used per token in the forward pass, as reported by the publisher.
- Active non-embedding (stated)
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.
- Counted total params
71.71B parameters
Calculated: embedding + output head + all experts + other weights, swiss-ai/Apertus-v1.5-70B-base@0f3fbdd safetensors headers; reproduces stated total (backbone) and active (with output head)
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.
- Counted embedding + head
3.259B parameters
Calculated: input embedding + output head, swiss-ai/Apertus-v1.5-70B-base@0f3fbdd safetensors headers
Input embedding plus output head (once if tied), counted from the checkpoint.
- Non-emb. params (counted)
68.45B parameters
Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, swiss-ai/Apertus-v1.5-70B-base@0f3fbdd safetensors headers; reproduces stated total (backbone) and active (with output head)
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.
- Active non-emb. (counted)
68.45B parameters
Calculated: all weights except embedding and output head, swiss-ai/Apertus-v1.5-70B-base@0f3fbdd safetensors headers; reproduces stated total (backbone) and active (with output head)
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.
- Counting scope
70B is a size designation; the checkpoint's 71,942,010,689 parameters include the image and audio tokenizers and the 266,752-row input embedding.
Source: config · model.safetensors.index.json; metadata
"total_parameters": 71942010689
What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).
Shape
- Layers
80
Source: config · /text_config/num_hidden_layers
Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.
- d_model
8,192
Source: config · /text_config/hidden_size
Residual stream width.
- Vocab
266,752
Source: config · /text_config/vocab_size
Input embedding only: 131,072 text + 200 reserved + 131,072 image + 4,096 audio + 312 padding rows; the output head keeps the 131,072 text rows (/text_config/output_vocab_size).
Embedding table rows (vocab_size in config, may include padding).
- Tied embeddings
No
Source: config · /text_config/tie_word_embeddings
Input embedding and output head share weights.
- Tokenizer
byte-level BPE (Mistral-Nemo tekken, 131,072) extended with 200 reserved slots, 131,072 image codes (Emu3.5 vision tokenizer, one per 16×16 patch) and 4,096 audio codes (WavTokenizer)
Source: paper · p.9 §2.1 Vocabulary Extension
We extend the 131,072-entry vocabulary of Apertus 1.0 with 200 reserved multimodal slots, including structure markers, followed by 131,072 image codes
Text tokenizer adapted from Mistral-Nemo-Base-2407 per the Apertus 1.0 report §2.2.
Tokenizer type and lineage as described by the publisher (e.g. byte-level BPE shared with an earlier model).
- d_model / layers
102.4
Calculated: d_model / layers
Width-to-depth ratio, d_model divided by the number of blocks.
Attention
- Token mixers
GQA
Source: paper · p.9 §2.1 Model Architecture
such as grouped-query attention (GQA), RoPE, and RMSNorm
Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).
- Layer layout
80 GQA
Source: paper · p.9 §2.1
Apertus 70B, with 80 layers and 64 parallel attention heads.
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.
- Global-attention layers
80
Source: config · /text_config/num_hidden_layers
Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.
- Query heads
64
Source: config · /text_config/num_attention_heads
Query heads in the global-attention layers.
- KV heads
8
Source: config · /text_config/num_key_value_heads
Key/value heads in the global-attention layers. Not applicable to MLA, whose cache is a shared latent.
- Heads × head dim
8,192
Calculated: heads × head_dim
attn.heads × attn.head_dim, the query projection width of global layers.
- Q heads / KV head
8
Calculated: heads / kv_heads
attn.heads divided by attn.kv_heads in global layers.
- QK head dim
128
Inferred from the source; see the explanation below.
Source: code · modeling_apertus1p5.py; Apertus1p5TextAttention.__init__
self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
text_config has no head_dim, so hidden_size 8192 / 64 heads = 128.
Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).
- Positional encoding
RoPE (Llama-3-style frequency scaling)
Source: paper · p.9 §2.1 Rotary Positional Embeddings
following the LLaMA-3 implementation
Released config: rope_type llama3, factor 32, low_freq_factor 1, high_freq_factor 4, original_max_position_embeddings 8192.
Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.
- RoPE base
4,000,000
Source: config · /text_config/rope_parameters/rope_theta
500,000 during Apertus 1.0 pretraining and Apertus 1.5 stages 1–3; 2M at 32K, 4M from 65K.
RoPE base frequency of the global layers at release.
- QK norm
Yes
Source: code · modeling_apertus1p5.py; Apertus1p5TextAttention.__init__
self.q_norm = Apertus1p5TextRMSNorm(self.head_dim, config.rms_norm_eps)
Per-head RMSNorm on Q and K before RoPE; config qk_norm true, code builds the norms unconditionally.
Queries and keys are normalized per head before the dot product.
- QKV bias
No
Source: config · /text_config/attention_bias
Query/key/value projections have bias terms.
Feed-forward
- Activation
xIELU (non-gated; trainable α_p, α_n per layer)
Source: code · modeling_apertus1p5.py; Apertus1p5TextMLP.forward
return self.down_proj(self.act_fn(self.up_proj(x)))
Config hidden_act xielu. xIELU(x) = α_p·x² + 0.5x for x > 0, α_n(eˣ − 1) − α_n·x + 0.5x otherwise (Apertus 1.0 report §2.1).
FFN activation (SwiGLU, GeGLU, ...).
- Dense FFN dim
43,008
Source: config · /text_config/intermediate_size
Intermediate width of the dense FFN layers that are actually executed.
- FFN bias
No
Source: config · /text_config/mlp_bias
FFN projections have bias terms.
- Dense FFN / d_model
5.25
Calculated: ffn.dim / d_model
ffn.dim divided by shape.d_model.
Multi-token prediction
- MTP depth
Not disclosed
Additional future tokens predicted during pretraining (0 when the model was trained without MTP).
Normalization and stability
- Norm
RMSNorm
Source: code · modeling_apertus1p5.py; Apertus1p5TextDecoderLayer.__init__
self.attention_layernorm = Apertus1p5TextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
Normalization layer type in the blocks.
- Norm placement
pre
Source: code · modeling_apertus1p5.py; Apertus1p5TextDecoderLayer.forward
hidden_states = self.attention_layernorm(hidden_states)
Pre-norm, post-norm, or sandwich (pre and post).
- Norm epsilon
1e-05
Source: config · /text_config/rms_norm_eps
Epsilon of the block normalization layers.
Context length
- Pretraining context
4.096K tokens
Source: paper · p.10 §2.1 Context length
Both Apertus 8B and Apertus 70B were trained with a context of 4,096 tokens (about 3,000 words) during pretraining.
Apertus 1.5 continued pretraining stages 1–3 use 8,192.
Sequence length of the main pretraining stage.
- Trained context
262.1K tokens
Source: paper · p.23 §3.3
from 8,192 to 32,768, 65,536, 131,072, and finally 262,144 tokens.
Longest sequence length trained on after context extension.
- Advertised context
262.1K tokens
Source: model-card · README.md; intro
handle contexts up to 262,144 tokens
Config max_position_embeddings 262144.
Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.
- Extension method
8K → 32K → 65K → 131K → 262K after the stage-3 cooldown, 58.7B, 58.7B, 50.3B, 50.3B tokens; constant LR 2e-6, fresh optimizer per step; RoPE base 2M then 4M, Llama-3 scaling factor 8, 8, 16, 32; context parallelism 4, 8, 16, 32
Source: paper · p.107 App. G
The RoPE base grows to 2M for the first step and 4M thereafter, with scaling factors of 8, 8, 16, and 32.
Apertus 1.5 continued pretraining resumes from the Apertus 1.0 checkpoint after its cooldown, at RoPE base 500,000 and factor 8; the Apertus 1.0 4K → 64K extension is not in this lineage.
How context was extended (stages, RoPE base change, YaRN factor).
Pretraining recipe
- Pretraining tokens
17.23T tokens
Inferred from the source; see the explanation below.
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2
trained from scratch on 15T tokens of openly documented text
Sum of train.stages: 15T Apertus 1.0 + 2,014B Apertus 1.5 stages 1–3 (Table 3: 403B + 1,208B + 403B) + 218.1B long-context (Table F.2: 58.72B + 58.72B + 50.33B + 50.33B).
Tokens seen in pretraining including annealing and context extension, as reported.
- Optimizer
AdEMAMix (α 8, β3 0.9999, α and β3 warmup over 100,000 iterations)
Source: paper · p.107 App. G
Both models train with AdEMAMix (β 1 = 0.9, β 2 = 0.999, α = 8, β 3 = 0.9999, with the β 3 and α warmup over 100,000 iterations)
Same settings in Apertus 1.0 pretraining (Apertus 1.0 report Table C.4); optimizer state reset at the start of continued pretraining.
Optimizer used for the bulk of parameters.
- Peak LR
1e-05
Source: paper · p.11 Table 2
Apertus 70B AdEMAMix 4096 8.4M → 16.8M 1.1M 1.0e-5 15T
Apertus 1.0 pretraining; Apertus 1.5 continued pretraining re-warms to 2e-5.
Peak learning rate of the main pretraining stage.
- Final LR
2e-06
Source: paper · p.107 Table G.1
Minimum learning rate 1.1 × 10−5 2 × 10−6
End of the Apertus 1.5 stage-3 cooldown, held through context extension; Apertus 1.0 cooled down to 1e-6.
Learning rate at the end of the decay.
- LR schedule
Apertus 1.0: WSD, warmup from 0.1× peak, constant 1e-5, 1-sqrt cooldown to 1e-6 over 13.5T–15T; Apertus 1.5: re-warm 0 → 2e-5, constant through stages 1–2, 1-sqrt decay to 2e-6 over stage 3, constant 2e-6 in context extension
Source: paper · p.20 §3.2
the learning rate then stays at its peak through the first two stages and decays to 10% of the peak with a 1 − sqrt decay over the third.
Apertus 1.0 schedule: Apertus 1.0 report §2.3 Learning Rate Schedule and Cooldown.
Learning-rate schedule shape with phase boundaries.
- Warmup
16.78B tokens from 0.1× peak (Apertus 1.0); 250 iterations, 4.2B tokens, from 0 (Apertus 1.5)
Source: paper · p.12 §2.3; p.82 Table C.4
Our LR warmup for both models starts from 0.1 the peak LR and is linearly increased for 16.8B tokens.
Table C.4: 16.78BT. Apertus 1.5 warmup from Table G.1 and §3.2.
Warmup length in steps or tokens.
- Batch size
Apertus 1.0: 2,048 × 4,096 (8.4M tokens), doubled to 16.8M tokens at 4.4T; Apertus 1.5: 2,048 × 8,192 (16.8M tokens), held constant through context extension
Source: paper · p.12 §2.3 Batch Size and Sequence Length
After 8T tokens for the 8B model and 4.4T for the 70B, we intentionally doubled both the number of nodes and the batch size at this stage, while keeping the learning rate unchanged.
Apertus 1.5 values: Table G.1 global batch 2,048 at 8,192 and §3.3 (sequences halved per doubling of length).
Global batch size, including any batch-size ramp: sequences × sequence length (token total), or the token total alone when only that is stated.
- Batch (tokens)
16.78M tokens
Inferred from the source; see the explanation below.
Source: paper · p.12 §2.3 Batch Size and Sequence Length
an initial batch size of 1024 (4.2M tokens) and 2048 (8.4M tokens) for the 8B and 70B models, respectively
Apertus 1.0 after the doubling at 4.4T: 2 × 2,048 × 4,096; Apertus 1.5 keeps it (2,048 × 8,192).
Global batch size in tokens during the main pretraining stage, after any ramp (sequences × sequence length when stated in sequences, marked inferred).
- Weight decay
0.1
Source: paper · p.107 App. G
gradient clipping at 0.1, and weight decay of 0.1, applied also to the xIELU parameters and the normalization gains.
Decoupled weight decay coefficient.
- Adam betas
(0.9, 0.999)
Source: paper · p.107 App. G
Both models train with AdEMAMix (β 1 = 0.9, β 2 = 0.999, α = 8, β 3 = 0.9999
(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.
- Grad clip
0.1
Source: paper · p.15 §2.6 Gradient Clipping
Our experiments led to set a clipping value of 0.1.
Clipping is active at almost every step; also 0.1 in continued pretraining (Apertus 1.5 report App. G).
Global gradient-norm clipping threshold.
- Training precision
BF16 (Apertus 1.0 pretraining; an FP8 trial on the 8B was rolled back)
Source: paper · p.16 §2.6 FP8 Training
We therefore decided to roll back and continue with the BF16 training normally.
The Apertus 1.5 report does not state the continued-pretraining precision.
Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.
- Initialization
Apertus 1.0: std 0.008944; Apertus 1.5: new image/audio embedding and output rows random at ~1/10 the norm of trained text rows
Source: paper · p.82 Table C.4
Initialization std 0.008944
Table C.4 lists one value without a per-size split. Row init from the Apertus 1.5 report §3.2.
Weight initialization (std or scheme) as stated by the publisher; config initializer_range alone is not evidence.
- Stability measures
QK-norm; grad clip 0.1; weight decay on norm gains and xIELU parameters
Source: paper · p.21 §3.2
including normalization gains and the trainable parameters of the xIELU activation (Huang & Schlag, 2024), which we find helps reduce large activation and activation-gradient norms.
Decay on gains and xIELU parameters is stated for continued pretraining; QK-norm and clip from the Apertus 1.0 report §2.1, §2.6.
Interventions used during pretraining to keep it stable (QK-clip, z-loss, logit caps, spike handling), with coefficients.
- Document masking
cross-document attention masking; position ids reset at document boundaries
Source: paper · p.107 App. G
All stages pack documents with cross-document attention masking: position identifiers reset at document boundaries, and end-of-document positions are excluded from the loss.
Apertus 1.0 pretraining also masks cross-document attention (Apertus 1.0 report §2.1).
Whether attention crosses document boundaries within a packed sequence.
- Extra objectives
Goldfish loss (k = 50, h = 50: ~2% of targets dropped by hashed 50-token context); image and audio tokens are targets only in Apertus 1.5 stage 1
Source: paper · p.21 §3.2
We keep the Goldfish loss of Apertus 1.0 with its parameters unchanged
k and h from Apertus 1.0 report Table C.4; ~0.3T of the 15T masked.
Objectives beyond next-token prediction and MTP (fill-in-the-middle rate, distillation, multimodal tokens from the start).
- Data mix
Apertus 1.0: 15T text in 1811 languages, robots.txt opt-outs applied retroactively, five data phases with rising math/code share (DCLM-Edu and FineWeb-2-HQ top 10% from 12T); Apertus 1.5 stages 1–3: 668B vision, 234B audio (both incl. paired text), 1,111B text-only tokens
Source: paper · p.17 §3; p.22 §3.3
Our data contains 1811 languages (1868 language-script pairs)
Apertus 1.5 per-stage modality split in its Table 3 and text mix in its Table 4.
Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).
- Hardware
NVIDIA GH200 on Alps (CSCS); Apertus 1.0 70B on 2048–4096 GPUs over several months; Apertus 1.5 on 512 nodes (2048 GPUs)
Source: paper · p.43 §6.1.2
running, for the 70B model, at scales from 2048 to 4096 GPUs over several months.
Apertus 1.5 node count from its Table G.1 and §6.1 (four GH200 per node).
Accelerators used for pretraining, count and duration where stated.
- Accelerators
2,048–4,096 GH200
Source: paper · p.43 §6.1.2
running, for the 70B model, at scales from 2048 to 4096 GPUs over several months.
Apertus 1.0 run; chip from the same report §6.1 (Alps GH200); Apertus 1.5 continued pretraining used 512 nodes (2,048 GPUs).
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").
- Parallelism
TP4 · PP8 · VPP2
Source: paper · p.107 Table G.1
TP4, PP8, VPP2
Apertus 1.5 continued pretraining; §3.2 states it reuses the Apertus 1.0 tensor and pipeline parallelism.
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").
- Compute
6.74e24 FLOPs and ~6M GPU-hours for the 15T Apertus 1.0 run; ~1.50M GPU-hours (logged segments) for Apertus 1.5 stages 1–3
Source: paper · p.44 §6.2
the main run has consumed around 6 million GPU hours
FLOPs from Apertus 1.0 report App. E; continued-pretraining GPU-hours from Apertus 1.5 report §6.3, excluding context extension.
Pretraining compute as stated (GPU-hours or FLOPs).
- Stages
Apertus 1.0 pretraining
- Tokens
- 15000000000000
- Seq len
- 4096
- Lr
- 1e-05
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2; p.107 App. G
trained from scratch on 15T tokens of openly documented text spanning over 1000 languages
From scratch; five data phases; batch 8.4M → 16.8M tokens at 4.4T; 1-sqrt cooldown 13.5T–15T to 1e-6.
Apertus 1.5 stage 1: unpaired multimodal
- Tokens
- 403000000000
- Seq len
- 8192
- Lr
- 2e-05
- Batch sequences
- 2048
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2; p.107 App. G
1 49B (12.2%) 52B (13.0%) 301B (74.8%) 403B
Text 74.8%, image 12.2%, audio 13.0%; loss on all modalities; fresh optimizer state, re-warmup over 250 iterations.
Apertus 1.5 stage 2: paired multimodal
- Tokens
- 1208000000000
- Seq len
- 8192
- Lr
- 2e-05
- Batch sequences
- 2048
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2; p.107 App. G
2 512B (42.4%) 90B (7.4%) 606B (50.1%) 1,208B
Image-text and audio-transcript pairs plus text; loss on text tokens only.
Apertus 1.5 stage 3: cooldown
- Tokens
- 403000000000
- Seq len
- 8192
- Lr
- 2e-05
- Batch sequences
- 2048
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2; p.107 App. G
3 107B (26.6%) 92B (22.7%) 204B (50.7%) 403B
Higher-quality data, ~3B instruction-formatted tokens; 1-sqrt decay 2e-5 → 2e-6.
Apertus 1.5 long-context 32K
- Tokens
- 58720000000
- Seq len
- 32768
- Lr
- 2e-06
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2; p.107 App. G
Total 58,720,000,000 58,720,000,000 50,330,000,000 50,330,000,000
RoPE base 2M, factor 8; context parallelism 4.
Apertus 1.5 long-context 65K
- Tokens
- 58720000000
- Seq len
- 65536
- Lr
- 2e-06
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2; p.107 App. G
Total 58,720,000,000 58,720,000,000 50,330,000,000 50,330,000,000
RoPE base 4M, factor 8; context parallelism 8.
Apertus 1.5 long-context 131K
- Tokens
- 50330000000
- Seq len
- 131072
- Lr
- 2e-06
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2; p.107 App. G
Total 58,720,000,000 58,720,000,000 50,330,000,000 50,330,000,000
RoPE base 4M, factor 16; context parallelism 16.
Apertus 1.5 long-context 262K
- Tokens
- 50330000000
- Seq len
- 262144
- Lr
- 2e-06
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2; p.107 App. G
Total 58,720,000,000 58,720,000,000 50,330,000,000 50,330,000,000
RoPE base 4M, factor 32; context parallelism 32.
Source: paper · p.6 §1; p.22 Table 3; p.106 Table F.2; p.107 App. G
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.
Scaling & HP transfer
- HP selection rule
peak LR 2e-5 = 2× Apertus 1.0 70B peak, chosen by ablation
Source: paper · p.20 §3.2
For the 70B model, we double the peak of Apertus 1.0 70B to 2 × 10−5 , which outperforms the original value in ablations
Apertus 1.0 set no LR scaling rule (report_v1 p.16 §2.5).
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.
- Scaling ladder
1.5B per-change ablations at 100B tokens, merged into a 3B run at 100B tokens
Source: paper · p.13 §2.4 Table 3
we run a separate ablation experiment on a 1.5B model scale with 100B tokens of our main datamix. The baseline is a standard Llama-style decoder with AdamW and a tuned cosine learning rate schedule. After verification, we merge all successful changes into a 3B model with 100B tokens
Apertus 1.0 architecture and optimizer ablations; Apertus 1.5 continues the 1.0 70B checkpoint.
Smaller models trained to choose the architecture or recipe (sizes, token budgets), as disclosed.
- Tokens / param
246.173
Calculated: train.tokens / params.total
train.tokens divided by params.total.
- Tokens / active param (non-embedding)
251.739
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.
KV cache
- KV bytes / token
327,680 bytes
Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 80 × (2048 × 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.
- Global KV bytes / token
327,680 bytes
Calculated: 327680 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 80 × (2048 × 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.
- KV bytes / token / B active
4,681 bytes
Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 80 × (2048 × 16/8) B/entry
kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).
- KV at context / B active
613,566,757 bytes
Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 80 × (2048 × 16/8) B/entry
kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).
- KV at context
42,949,672,960 bytes
Calculated: 327680 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 80 × (2048 × 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.