stepfun-ai · vision language

Step-3.7-Flash

stepfun-ai/Step-3.7-Flash

Step-3.7-Flash at Q4_K_M is exactly 121,621,146,016 bytes (113.27 GiB / 121.62 GB) — an effective 4.832 bits per weight, not the nominal 4. Its KV cache at 32K is 13.03 GiB, not the 45.00 GiB a flat formula predicts.

From the file· summed from 4 file(s)From the file· KV per layer
Parameters
201B
Architecture
step35
45 layers
Context
262,144
native (config.json)
License
apache-2.0

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
IQ1_S40.03 GiB42,985,858,0161.708—bartowski
IQ1_M44.41 GiB47,681,234,9121.894—bartowski
IQ2_XXS2 shards51.22 GiB54,995,682,4962.185—bartowski
UD-IQ1_M3 shards52.85 GiB56,750,287,3282.255—unsloth
IQ2_XS2 shards56.76 GiB60,941,612,1922.421—bartowski
UD-IQ2_XXS3 shards57.47 GiB61,704,808,9282.451—unsloth
UD-IQ2_M3 shards57.58 GiB61,821,856,2242.456—unsloth
IQ2_S2 shards57.93 GiB62,196,901,0242.471—bartowski
IQ2_M2 shards63.68 GiB68,378,760,3842.717—bartowski
Q2_K2 shards66.34 GiB71,226,835,1042.830—bartowski
Q2_K_L2 shards66.82 GiB71,742,419,1362.850—bartowski
UD-IQ3_XXS3 shards68.54 GiB73,591,039,4882.924—unsloth
IQ3_XXS2 shards70.56 GiB75,758,680,9603.010—stepfun-ai
UD-IQ3_S3 shards74.54 GiB80,031,917,5683.180—unsloth
IQ3_XXS3 shards78.38 GiB84,162,180,3843.344—bartowski
Q3_K_S3 shards81.55 GiB87,566,150,9763.479—bartowski
UD-Q3_K_M3 shards83.13 GiB89,262,719,4883.546—unsloth
IQ3_XS3 shards85.44 GiB91,736,206,6243.645—bartowski
Q3_K_M3 shards85.51 GiB91,813,178,6563.648—bartowski
Q3_K_M3 shards87.36 GiB93,801,444,3523.727—stepfun-ai
UD-IQ4_XS3 shards88.79 GiB95,336,010,2083.788—unsloth
Q3_K_L3 shards88.94 GiB95,502,724,4163.794—bartowski
IQ3_M3 shards89.38 GiB95,976,156,4803.813—bartowski
UD-IQ4_NL3 shards90.63 GiB97,317,818,8803.866—unsloth
Q3_K_L3 shards95.46 GiB102,504,428,5444.072—stepfun-ai
IQ4_XS3 shards97.78 GiB104,993,562,6244.171—stepfun-ai
IQ4_XS3 shards99.94 GiB107,305,015,5524.263—bartowski
Q4_K_S3 shards103.84 GiB111,499,087,8724.430—stepfun-ai
IQ4_NL3 shards105.54 GiB113,325,345,0884.502—bartowski
Q4_03 shards105.79 GiB113,589,193,0244.513—bartowski
UD-Q4_K_S4 shards106.32 GiB114,163,192,4484.536—unsloth
Q4_K_S3 shards109.05 GiB117,089,601,8244.652—bartowski
Q4_K_M4 shards113.27 GiB121,621,146,0164.832—bartowski
Q4_K_L4 shards113.63 GiB122,012,989,8564.847—bartowski
UD-Q4_K_M4 shards113.71 GiB122,090,426,9764.851—unsloth
Q4_14 shards116.67 GiB125,278,316,9924.977—bartowski
Q5_K_S4 shards128.01 GiB137,450,531,2325.461—bartowski
UD-Q5_K_S4 shards128.59 GiB138,072,760,9605.486—unsloth
Q5_K_M4 shards132.28 GiB142,034,323,9045.643—bartowski
Q5_K_L4 shards132.58 GiB142,360,172,9925.656—bartowski

KV cache by context

computed per layer — this model uses sliding-window attention
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0962.53 GiB5.63 GiB2.22×12 / 33 / 0
8,1924.03 GiB11.25 GiB2.79×12 / 33 / 0
16,3847.03 GiB22.50 GiB3.20×12 / 33 / 0
32,76813.03 GiB45.00 GiB3.45×12 / 33 / 0
65,53625.03 GiB90.00 GiB3.60×12 / 33 / 0
131,07249.03 GiB180.00 GiB3.67×12 / 33 / 0

33 of 45 layers cache only a 512-token window rather than the full context, on a period of . Figures assume the default configuration; --swa-full disables the saving entirely.

Compare with

same modality, comparable size
vs Qwen3.6-35B-A3Bvs Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPvs Qwen3.5-9Bvs gemma-4-26B-A4B-itvs gemma-4-12B-itvs Qwythos-9B-Claude-Mythos-5-1MAll alternatives →Will it fit on my card? →

Will it run on your card?

full quant x context sweep
By VRAM:8GB12GB16GB24GB32GB48GB
Radeon RX 6500 XT 4GBGeForce RTX 3050 6GBGeForce RTX 5050 8GBGeForce RTX 3080 10GBGeForce RTX 2080 Ti 11GBGeForce RTX 5070 12GBGeForce RTX 5060 Ti 16GBApple M3 Pro 18GBGeForce RTX 3080 Ti 20GBGeForce RTX 5090 D V2 24GBGeForce RTX 5090 D 32GBApple M5 Max 36GBApple M5 Pro 48GBApple M5 Max 64GBApple M3 Ultra 96GBApple M5 Max 128GBApple M2 Ultra 192GBApple M3 Ultra 256GBApple M3 Ultra 512GB

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 105.49 GiB. The real file is 113.27 GiB, because a quantization is a mixture and some tensors are always kept at higher precision. The larger discrepancy is the cache: a flat formula gives 45.00 GiB at 32K context where the real figure is 13.03 GiB, because most of this model's layers cache a fixed window rather than the whole context.

Architecture

from config.json
Layers
45
Attention heads
64
KV heads
64
Head dim
128
Hidden size
4096
Vocab
128,896
Sliding window
512
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
—

Questions people ask

How much VRAM does Step-3.7-Flash need?
Q4_K_M is exactly 121,621,146,016 bytes (113.27 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Step-3.7-Flash's KV cache?
13.03 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Which quantization of Step-3.7-Flash should I use?
Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.