Research Vault

StepFun Step 5 Preview

Properties1
tagshuggingface, coding, agentic, moe, vision-language, long-context, fp8, instruct, finance

3 min read

StepFun Step 5 Preview

One-line: StepFun’s flagship agentic model — 600B-total / 27B-active MoE with 1M context and vision input, positioned for software engineering and finance; open weights released 2026-10-15 (API since Sep 20; NVIDIA publicly congratulated the drop as StepFun’s “largest open weights release”). ⚠ Benchmark chart is image-only and lab-reported; ⚠ 650.3 GB FP8 across 22 shards — data-center serving only (4-8 GPU vLLM/SGLang recipes), no local path at release.

⚠ All numbers below come from the card’s single image chart (read/verified via vision extraction, two passes); no SWE-bench Verified / LiveCodeBench / BFCL / τ-Bench numbers published. GDPval-AA v2.1 row is sourced by StepFun from Artificial Analysis (as of 2026-09-20). Cross-model comparisons shown: Kimi K3 (Max), GLM-5.3 (Max), GPT-6 Astra (Max), Claude Opus 5 (Max).

Coding benchmarks

BenchmarkScoreSource
HumanEval—not published
HumanEval+—not published
MBPP—not published
LiveCodeBench—not published
Aider polyglot—not published
SWE-bench Verified—not published (DeepSWE v1.1 used instead)
SWE-bench Multilingual—not published
DeepSWE v1.167.7lab — card chart (SWE-agent harness); vs Kimi K3 67.5, GLM-5.3 66.9, GPT-6 Astra 74.1, Claude Opus 5 74.0
StepCodeBench (in-house)49.0lab — card chart; vs Kimi K3 43.9, GLM-5.3 40.2, GPT-6 Astra 61.0, Claude Opus 5 63.9
ProgramBench80.5lab — card chart; vs Kimi K3 77.8, GLM-5.3 72.0, GPT-6 Astra 85.4, Claude Opus 5 82.3
Terminal-Bench v433.3lab — card chart; vs Kimi K3 12.6, GLM-5.3 41.9, GPT-6 Astra 57.9, Claude Opus 5 52.3

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL—not published
τ-Bench—not published
ToolACE—not published
GAIA—not published
Agents’ Last Exam (ALE-CLI)29.5lab — card chart; vs Kimi K3 27.6, GLM-5.3 28.6, Claude Opus 5 28.6, GPT-6 Astra 33.3
GDPval-AA v2.11566lab — card chart, from Artificial Analysis as of 2026-09-20; vs GLM-5.3 1645, Claude Opus 5 1708, GPT-6 Astra 1542, Kimi K3 1524
FrontierFinance66.4lab — card chart; vs Claude Opus 5 69.7, GLM-5.3 64.1, Kimi K3 62.6
DRACO (deep research)83.3lab — card chart; vs Claude Opus 5 87.6, GLM-5.3 82.3, Kimi K3 78.5

Reasoning (context only)

BenchmarkScoreSource
MMLU / MMLU-Pro—not published
GPQA Diamond—not published
MATH—not published
AIME 2024/2025—not published
HLE—not published

Efficiency

  • Active params: 27B per token of 600B total (sparse MoE; expert count not disclosed)
  • Context length: 1,048,576 (1M); card claims max output up to 1M (third-party pages cite 64k practical output)
  • VRAM: FP8 main checkpoint = 650.3 GB across 22 shards (verified; main repo IS the FP8 release); BF16 and NVFP4 siblings on HF + ModelScope. Serving recipes: vLLM TEP4 (4 GPUs, --spec-tokens 3, tool parser step5, max-model-len 1048576) or DEP4; SGLang TP8 + EAGLE spec decode. ⚠ Not a local model — 4-8 GPU-class; watch for community int4/GGUF quants.
  • Release mechanics: repos/weights staged since 2026-09-20 (API launch), made public with the open-weights release 2026-10-15 (README/final card commits 10-13 → 10-15 01:28Z); 144 downloads / 63 likes at note time (launch day).

What makes it notable

StepFun’s biggest open release and a credible open challenger in the “long-horizon agentic work” class: 600B/27B MoE at 1M context with vision, roughly matching Kimi K3 on DeepSWE (67.7 vs 67.5) and beating GLM-5.3 on StepCodeBench / ProgramBench / DRACO rows, while trailing both on Terminal-Bench v4 (33.3 vs 41.9 vs Claude Opus 5 52.3). Its differentiation is finance (FrontierFinance 66.4, “particular strength in finance”) and cheap API pricing (2.70 per M in/out) as the hosted preview. For this user’s coding/agent loop: no SWE-bench Verified, no BFCL — the published suite is custom-heavy (DeepSWE, StepCodeBench, ProgramBench) and 650 GB FP8 rules out local runs; interesting mainly as an OpenRouter-hosted agent backend and as a signal for the 600B-class open field. Watch for: SWE-bench Verified / BFCL numbers, community quants, and independent (AA) full re-evaluation of the open checkpoint.

See also

Graph View