Step 5 Preview step-5-preview

stepfun-ai · stepfun-ai

unranked — no quality signals yet

params
604.3B
arch
moe
context
1024k
license
other open weights
released
Sep 2026
downloads/30d
3.1k

# architecture

attn shared MHA ×64 router top-8 of 352 expert ×352 expert out ×92 layers ctx 1,048,576 604.3B pool
layers
92
d_model
4096
heads
64q / 64kv
head dim
192
experts
top-8 of 352
vocab
128896
family
step5

signal path of one layer, generated from the registry's structured fields — dimension lines quote the model's real numbers; an MoE trace forks at the router, a dense trace runs straight through. Geometry fields come from the repo's config.json.

# trend

collecting history — trends appear after a few daily runs.

# where to run

providerquantctx$/M in$/M out$/M cacheprice srctpsuptime✓
DeepInfrafp8 1024k $1.00$2.70— deepinfra — 100.0% ✓
Featherlessfp4 256k $1.00$2.70$0.10 via openrouter — 100.0%
InferenceNetfp4 1024k $1.00$2.70$0.05 via openrouter — 100.0%
InferenceNetfp4 1024k $1.00$2.70$0.05 via openrouter — 100.0%
Novita 1000k $1.00$2.70— novita — — ✓
Opencode Go 1000k $1.00$2.70— via litellm — —
StepFunfp8 1000k $1.00$2.70$0.10 via openrouter — 100.0%

sorted by blended price ((3·input + output) / 4 per 1M) · ✓ = the provider's own catalog confirms the offer · "via …" prices are what the aggregator routing the offer charges, not the provider's own list price

source aliases
epoch
Step 5 Preview
openrouter
stepfun/step-5-preview