Step-3.5-Flash-Base

2026-02-11

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.

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.

text

Inferred from the source; see the explanation below.

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").

196B parameters

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

11B parameters

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

11B parameters

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.

197B parameters

Calculated: embedding + output head + all experts + other weights, stepfun-ai/Step-3.5-Flash-Base@19623ae 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.

1.055B parameters

Calculated: input embedding + output head, stepfun-ai/Step-3.5-Flash-Base@19623ae safetensors headers

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

195.9B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, stepfun-ai/Step-3.5-Flash-Base@19623ae 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.

4.656B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, stepfun-ai/Step-3.5-Flash-Base@19623ae 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.

10.93B parameters

Calculated: non-expert weights + routed experts × 8/288, excluding embedding and output head, stepfun-ai/Step-3.5-Flash-Base@19623ae 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.

Backbone only: 198B total / 13B active with the 3 MTP heads; active count excludes embedding and output matrices.

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

45

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

4,096

Residual stream width.

128,815

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

No

Input embedding and output head share weights.

91.0222

Calculated: d_model / layers

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

GQA, SWA

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

3 SWA : 1 global (leading global layer + 11 × [3 SWA, 1 global])

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.

12

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

2.75

Calculated: attn.local_layers / attn.global_layers

attn.local_layers divided by attn.global_layers, the sliding-window to global interleave ratio (5 for '5 SWA : 1 global').

33

Layers with a fixed attention window.

64

Query heads in the global-attention layers.

8

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

8,192

Calculated: heads × head_dim

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

8

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).

512 tokens

Sliding-window size of the local layers.

8

KV heads in sliding-window layers (stated even when equal to attn.kv_heads).

96

Query heads in sliding-window layers (stated even when equal to attn.heads).

128

Query/key head dimension in sliding-window layers. Value head dimensions may differ; see attn.local_v_head_dim.

partial RoPE (global, 64 of 128 dims), RoPE (SWA, all 128 dims)

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

1,000,000

RoPE base frequency of the global layers at release.

10,000

RoPE base frequency of the sliding-window layers (stated even when equal to attn.rope_theta).

0.5

Fraction of each head's QK dimensions that are rotated.

Yes

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

head-wise scalar sigmoid gate on each head's output, before o_proj; all layers

Gate applied to the attention output before the output projection, with its placement.

No

Query/key/value projections have bias terms.

SwiGLU

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

11,264

Intermediate width of the dense FFN layers that are actually executed.

gate capped at 7, up clamped to [−7, 7] in routed experts of layers 43–44; shared expert of layer 44 limit 16

Clamping of GLU gate / up activations (e.g. SwiGLU limit 7), with the layers it applies to.

No

FFN projections have bias terms.

2.75

Calculated: ffn.dim / d_model

ffn.dim divided by shape.d_model.

288

Routed experts per MoE layer.

8

Routed experts selected per token.

1

Always-on experts per MoE layer.

1,280

Intermediate width of one routed expert.

1,280

Total intermediate width of the shared expert(s).

3

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

sigmoid

Gating function over expert logits (softmax, sigmoid).

aux-loss-free bias (update rate 0.001 for the first 14.6T, decayed to 0 during annealing) + EP-group balance loss 0.001

Load-balancing method, with coefficients where disclosed.

3

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

Yes

Selected expert weights are renormalized to sum to 1.

0.3125

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

36

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.0561224

Calculated: params.active / params.total

params.active divided by params.total.

1

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.

gated SWA attention + dense SwiGLU FFN; H-Norm(hidden) and E-Norm(next-token embedding) concatenated, then linear projection

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

dense

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

No

Inferred from the source; see the explanation below.

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

810M parameters

Parameters of all MTP modules as reported by the publisher.

0.3 in pre-training stage 1; 0.1 in stage 2

Weight of the MTP loss, including its schedule.

Yes

The released checkpoint contains the MTP module weights.

RMSNorm (1+w gain)

Normalization layer type in the blocks.

pre

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

1e-05

Epsilon of the block normalization layers.

4.096K tokens

Sequence length of the main pretraining stage.

32.77K tokens

Longest sequence length trained on after context extension.

262.1K tokens

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

last 1T tokens of annealing at 32K; full-attention RoPE θ 10K → 1M, SWA θ stays 10K

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

17.6T tokens

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

Muon (Polar Express orthogonalization, 6 iterations)

Optimizer used for the bulk of parameters.

0.00025

Peak learning rate of the main pretraining stage.

2e-05

Learning rate at the end of the decay.

2,000-step linear warmup; cosine 2.5e-4 → 5e-5 over stage 1 (14.6T); cosine 5e-5 → 2e-5 over the 2T 4K anneal; constant 2e-5 for the 1T 32K portion

Learning-rate schedule shape with phase boundaries.

2,000 steps

Warmup length in steps or tokens.

4,096 → 16,384 × 4,096 (67.1M tokens) over the first 400B tokens, then constant; 2,048 × 32,768 (67.1M tokens) in the 32K anneal

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

67.11M tokens

Inferred from the source; see the explanation below.

Global batch size in tokens during the main pretraining stage, after any ramp (sequences × sequence length when stated in sequences, marked inferred).

0.1

Decoupled weight decay coefficient.

Not disclosed

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

1

Global gradient-norm clipping threshold.

mixed precision; Muon Polar Express iteration (state and intermediates) in FP16

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

FP16 Polar Express in Muon; SwiGLU activation clipping in deep MoE layers; zero-centered RMSNorm; QK-norm; head-wise attention gate; per-expert activation/parameter-norm monitoring

Interventions used during pretraining to keep it stable (QK-clip, z-loss, logit caps, spike handling), with coefficients.

metadata prefix (content type, language, domain, source) on every sample; its loss masked after ~3.8T tokens; PR-Dialogue template scaffolding masked during annealing

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

StepCrawl web and ePub/PDF, code (OpenCoder filters relaxed to ≤6 violations), GitHub PR/issue/commit (incl. 90B PR-Dialogue tokens), math/STEM, synthetic tool-use and reasoning; annealing upweights code, PR/issue/commit and high-quality knowledge

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

4,096 NVIDIA H800; 8 GPUs per node, 8×200 Gbps RoCE

Accelerators used for pretraining, count and duration where stated.

4,096 H800

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").

PP8 · VPP · EP8 · ZeRO-1 DP

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").

• pre-training stage 1

  Tokens

  14600000000000

  Seq len

  4096

  Lr

  0.00025

  Batch sequences

  16384

  Source: paper · p.14 §4.2.2; p.15 §4.2.3

  Pre-training Stage 1: Open-domain pre-training (14.6T tokens, 4k context).

  Open-domain mix; batch ramps from 4,096 over the first 400B tokens; cosine to 5e-5; MTP weight 0.3.

• pre-training stage 2 (annealing, 4K)

  Tokens

  2000000000000

  Seq len

  4096

  Lr

  5e-05

  Batch sequences

  16384

  Source: paper · p.14 §4.2.2; p.15 §4.2.3

  This stage starts with 2T tokens at 4k context

  Annealed mix; cosine 5e-5 → 2e-5; MTP weight 0.1; expert-bias update rate decays to 0.

• pre-training stage 2 (annealing, 32K)

  Tokens

  1000000000000

  Seq len

  32768

  Lr

  2e-05

  Batch sequences

  2048

  Source: paper · p.14 §4.2.2; p.15 §4.2.3

  then transitions to 1T tokens at 32k context under the same annealed mixture to initialize long-context training.

  Same annealed mix; constant LR; full-attention RoPE θ 1M.

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.

30B-A3B (48 layers, d 2048, 1.4T tokens) for attention-layout and data ablations; 100B-A10B (43 layers, d 4096, 250B tokens) for head-wise gate vs sink

Smaller models trained to choose the architecture or recipe (sizes, token budgets), as disclosed.

89.7959

Calculated: train.tokens / params.total

train.tokens divided by params.total.

1,610.04

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.

49,680 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (2048 × 16/8) B/entry; SWA KV: 33 × (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.

49,152 bytes

Calculated: 49152 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (2048 × 16/8) B/entry; SWA KV: 33 × (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.

4,516 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (2048 × 16/8) B/entry; SWA KV: 33 × (2048 × 16/8) B/entry

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

591,968,815 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (2048 × 16/8) B/entry; SWA KV: 33 × (2048 × 16/8) B/entry

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

6,511,656,960 bytes

Calculated: 49152 B × 131072 (Global KV) + 135168 B × min(131072, 512) (SWA KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (2048 × 16/8) B/entry; SWA KV: 33 × (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.

17

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).