Step-3
StepFun · reasoning checkpoint
No base checkpoint is public; facts describe the released post-trained checkpoint (stepfun-ai/step3, a VLM); architecture facts are for the language model.
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
2025-07-25
Source: paper · p.1 arXiv stamp
arXiv:2507.19427v1 [cs.LG] 25 Jul 2025
Report predates the weights (release blog dated July 31, 2025).
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; 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
Source: blog · Step3 blog; Pretrain, multimodal pathway
our 5B-parameter vision encoder extracts dense image features
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
321B parameters
Source: model-card · README.md; Introduction
with 321B total parameters and 38B active
Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.
- Active params
38B parameters
Source: model-card · README.md; Introduction
with 321B total parameters and 38B active
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
316.7B parameters
Calculated: embedding + output head + all experts + other weights, stepfun-ai/step3@e2d67fc safetensors headers; reproduces stated total (other modalities) 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
1.847B parameters
Calculated: input embedding + output head, stepfun-ai/step3@e2d67fc safetensors headers
Input embedding plus output head (once if tied), counted from the checkpoint.
- Non-emb. params (counted)
314.8B parameters
Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, stepfun-ai/step3@e2d67fc safetensors headers; reproduces stated total (other modalities) 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.
- Largest layer (counted)
5.565B parameters
Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, stepfun-ai/step3@e2d67fc 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.
- Active non-emb. (counted)
37.38B parameters
Calculated: non-expert weights + routed experts × 3/48, excluding embedding and output head, stepfun-ai/step3@e2d67fc safetensors headers; reproduces stated total (other modalities) 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
321B includes the 5B vision encoder; the LLM alone is 316B. 38B active per text token.
Source: model-card · README.md; Step3 model card table
Total Parameters (LLM) | 316B
What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).
Shape
- Layers
61
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
7,168
Source: config · /text_config/hidden_size
Residual stream width.
- Vocab
128,815
Source: config · /text_config/vocab_size
Embedding table rows (vocab_size in config, may include padding).
- Tied embeddings
No
Inferred from the source; see the explanation below.
Source: config · model.safetensors.index.json; weight_map
"lm_head.weight": "model-00005.safetensors"
Inferred from the checkpoint: lm_head.weight (shard 5) and model.embed_tokens.weight (shard 1) are separate [128815, 7168] tensors with different bytes; no config key or code states tying (config.json has no tie_word_embeddings).
Input embedding and output head share weights.
- Tokenizer
DeepSeek-V3 tokenizer
Source: model-card · README.md; Step3 model card table
Tokenizer | Deepseek V3
Tokenizer type and lineage as described by the publisher (e.g. byte-level BPE shared with an earlier model).
- d_model / layers
117.508
Calculated: d_model / layers
Width-to-depth ratio, d_model divided by the number of blocks.
Attention
- Token mixers
MFA
Source: model-card · README.md; Step3 model card table
Attention Mechanism | MFA
Multi-matrix Factorization Attention: query down-projected 7168 → 2048 (/text_config/share_q_dim), RMSNorm, up-projected to 64×256; one K and one V head shared by all query heads.
Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).
- Global-attention layers
61
Source: model-card · README.md; Step3 model card table
Number of Layers (Dense layer included)|61
Every layer uses full MFA attention (Step3vDecoderLayer).
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
1
Source: paper · p.2 §2
Key and a Value head, all with a dimension of 256.
Step3vAttention hardcodes num_key_value_heads = 1 (/text_config/num_attention_groups 1 is not read). The cache is 512 elements per token per layer, matching report Table 3 (1.02e9 bytes at 32K in FP8 = 32,768 × 61 × 512).
Key/value heads in the global-attention layers. Not applicable to MLA, whose cache is a shared latent.
- Heads × head dim
16,384
Calculated: heads × head_dim
attn.heads × attn.head_dim, the query projection width of global layers.
- Q heads / KV head
64
Calculated: heads / kv_heads
attn.heads divided by attn.kv_heads in global layers.
- QK head dim
256
Source: config · /text_config/head_dim
Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).
- Positional encoding
RoPE
Source: code · modeling_step3.py; Step3vAttention.forward
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
All 256 dims rotated; no rope_scaling in config.json.
Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.
- RoPE base
500,000
Source: config · /text_config/rope_theta
RoPE base frequency of the global layers at release.
- QK norm
No
Source: code · modeling_step3.py; Step3vAttention.forward
query_states = self.inter_norm(query_states)
The only attention norm is RMSNorm on the 2048-dim query latent before up-projection; per-head Q and K are not normalized.
Queries and keys are normalized per head before the dot product.
- QKV bias
No
Source: code · modeling_step3.py; Step3vAttention.__init__
self.k_proj = nn.Linear(config.hidden_size, self.head_dim, bias=False)
q_proj, wq, v_proj and o_proj are also bias-free.
Query/key/value projections have bias terms.
Feed-forward
- Activation
SwiGLU
Source: code · modeling_step3.py; Step3vMLP.forward
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
act_fn is SiLU; routed experts use the same form (Step3vMoEMLP.get_expert_output).
FFN activation (SwiGLU, GeGLU, ...).
- Dense FFN dim
18,432
Source: config · /text_config/intermediate_size
Dense FFN of layers 0–3 and 60.
Intermediate width of the dense FFN layers that are actually executed.
- FFN bias
No
Source: code · modeling_step3.py; Step3vMLP.__init__
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
FFN projections have bias terms.
- Dense FFN / d_model
2.57143
Calculated: ffn.dim / d_model
ffn.dim divided by shape.d_model.
Mixture of experts
- Routed experts
48
Source: config · /text_config/moe_num_experts
Routed experts per MoE layer.
- Experts per token
3
Source: config · /text_config/moe_top_k
Routed experts selected per token.
- Expert dim
5,120
Source: config · /text_config/moe_intermediate_size
Intermediate width of one routed expert.
- Leading dense layers
4
Source: config · config.json text_config.moe_layers_enum
"moe_layers_enum": "4,5,6,7,
moe_layers_enum lists layers 4–59, so the last layer (60) is dense too: 5 dense layers in total (card).
Number of initial blocks that use a dense FFN instead of MoE.
- Router score
softmax
Source: code · modeling_step3.py; Step3vMoEMLP.forward
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
Gating function over expert logits (softmax, sigmoid).
- Load balancing
Not disclosed
Load-balancing method, with coefficients where disclosed.
- Renormalize top-k
No
Source: code · modeling_step3.py; Step3vMoEMLP.forward
routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1) # we cast back to the input dtype
Top-k softmax weights are used without renormalization, consistent with /text_config/norm_expert_weight false (which the code does not read).
Selected expert weights are renormalized to sum to 1.
- Expert dim / d_model
0.714286
Calculated: moe.expert_dim / d_model
moe.expert_dim divided by shape.d_model.
- Sparsity (experts / top-k)
16
Calculated: experts / top_k
Routed experts divided by experts selected per token.
- Active / total
0.11838
Calculated: params.active / params.total
params.active divided by params.total.
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_step3.py; Step3vDecoderLayer.__init__
self.input_layernorm = Step3vRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
Normalization layer type in the blocks.
- Norm placement
pre
Source: code · modeling_step3.py; Step3vDecoderLayer.forward
hidden_states = self.input_layernorm(hidden_states)
Pre-norm, post-norm, or sandwich (pre and post).
- Norm epsilon
1e-05
Source: code · configuration_step3.py; Step3TextConfig.__init__
rms_norm_eps: float = 1e-5,
config.json sets no rms_norm_eps, so the Step3TextConfig default applies.
Epsilon of the block normalization layers.
Context length
- Pretraining context
Not disclosed
Sequence length of the main pretraining stage.
- Trained context
Not disclosed
Longest sequence length trained on after context extension.
- Advertised context
65.54K tokens
Source: model-card · README.md; Step3 model card table
Max Context Length | 65536
/text_config/max_seq_len and /text_config/max_position_embedding (RoPE cache length) are also 65536.
Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.
Pretraining recipe
- Pretraining tokens
>20T
Source: blog · Pretrain Data
During pretraining, Step3 processes over 20T text tokens across more than ten languages
Text tokens; 4T image-text mixed tokens are counted separately.
Tokens seen in pretraining including annealing and context extension, as reported.
- Optimizer
Not disclosed
Optimizer used for the bulk of parameters.
- Peak LR
Not disclosed
Peak learning rate of the main pretraining stage.
- Final LR
Not disclosed
Learning rate at the end of the decay.
- LR schedule
Not disclosed
Learning-rate schedule shape with phase boundaries.
- Warmup
Not disclosed
Warmup length in steps or tokens.
- Batch size
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.
- Weight decay
Not disclosed
Decoupled weight decay coefficient.
- Adam betas
Not disclosed
(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.
- Grad clip
Not disclosed
Global gradient-norm clipping threshold.
- Training precision
Not disclosed
Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.
- Extra objectives
multimodal: 4T image-text mixed tokens in pretraining; connector and LLM trained on 1.4T multimodal tokens with the 5B vision encoder frozen
Source: blog · Multimodal
In the second stage, the vision encoder is frozen, and the Connector and the LLM are trained on the full dataset, totaling 1.4T tokens.
Objectives beyond next-token prediction and MTP (fill-in-the-middle rate, distillation, multimodal tokens from the start).
- Data mix
>20T text tokens in >10 languages, 3.7T high-quality tokens reserved for annealing; web and licensed publisher data, MinHash-deduplicated; 4T image-text mixed tokens
Source: blog · Pretrain Data
of which 3.7T high-quality tokens are reserved for the annealing phases, while incorporating 4T image-text mixed tokens
Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).
- Stages
pretraining
Source: blog · Pretrain Data
During pretraining, Step3 processes over 20T text tokens
>20T text tokens in total, including annealing.
annealing
- Tokens
- 3700000000000
Source: blog · Pretrain Data
of which 3.7T high-quality tokens are reserved for the annealing phases
Several annealing phases on high-quality data.
Source: blog · Pretrain Data
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.
KV cache
- KV bytes / token
62,464 bytes
Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (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.
- Global KV bytes / token
62,464 bytes
Calculated: 62464 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (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.
- KV bytes / token / B active
1,644 bytes
Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (512 × 16/8) B/entry
kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).
- KV at context / B active
215,454,774 bytes
Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (512 × 16/8) B/entry
kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).
- KV at context
8,187,281,408 bytes
Calculated: 62464 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (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.