What are the strengths of this model?

1M-token context with native vision input Matches Kimi K3's AA index (44) at ~1/4 the parameter size Open weights planned for Oct 15, 2026 Aggressive API pricing at $1/$2.70 per 1M tokens

What are the weaknesses of this model?

Preview only; no independent evaluation published yet API-only until open weights ship on Oct 15 Sparse public benchmarks beyond the AA Intelligence Index StepFun's track record on open-weight delivery at this scale is unproven

What are the best use cases?

Long-context multimodal agents Cost-sensitive coding assistance Private deployment after the open-weight release

StepFunOpen Source

Step 5 Preview

Name: Step 5 Preview
Author: StepFun

Compare this model

StepFun's Step 5 Preview is a 600B-parameter sparse Mixture-of-Experts model previewed Sep 21, 2026, activating 27B parameters per token. It carries a 1M-token context window with native vision input and is priced at $1/$2.70 per million tokens. On the Artificial Analysis Intelligence Index it scores 44 — the same as Moonshot's Kimi K3, despite being roughly a quarter of its 2.8T size. StepFun plans to release open weights on Oct 15, 2026, which would make Step 5 one of the largest openly licensed models from a Chinese lab and a direct test of the 'API first, open weights later' release strategy.

Share:X Hatena LINE

Parameters

600B (MoE, 27B active)

Context Window

License

Apache 2.0 (open weights planned Oct 15)

Release Date

2026-09-21

Benchmark Performance

AA Intelligence Index

44.0

LMArena Elo

—

HLE

—

ARC-AGI-2

—

SWE-bench Verified

—

GPQA Diamond

—

MMLU-Pro

—

LiveCodeBench

—

AIME 2025

—

MATH-500

—

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

$2.7

Billing Mode: standard

Strengths

•1M-token context with native vision input
•Matches Kimi K3's AA index (44) at ~1/4 the parameter size
•Open weights planned for Oct 15, 2026
•Aggressive API pricing at $1/$2.70 per 1M tokens

Weaknesses

•Preview only; no independent evaluation published yet
•API-only until open weights ship on Oct 15
•Sparse public benchmarks beyond the AA Intelligence Index
•StepFun's track record on open-weight delivery at this scale is unproven

Use Cases

•Long-context multimodal agents
•Cost-sensitive coding assistance
•Private deployment after the open-weight release

Deep Analysis

Artificial Analysis Intelligence Index

Kimi K3와 동일, 규모는 약 4분의 1

Parameters

600B (MoE, 27B active)

희소 MoE

Context Window

1M tokens

네이티브 비전 입력

API Price

$1 / $2.70 per 1M tokens

표준 과금

Open Weights

Oct 15, 2026 (planned)

Apache 2.0 예정

Release

Sep 21, 2026 (preview)

StepFun

Strengths

・One of the largest openly licensed MoE planned from a Chinese lab at 600B scale
・1M-token context with native vision input for long-horizon multimodal agents
・Matches Kimi K3's AA index (44) at roughly a quarter of the parameter size
・Aggressive API pricing at $1/$2.70 per 1M tokens

Weaknesses

・Preview only; no independent evaluation has appeared yet
・API-only until open weights ship on Oct 15
・Sparse public benchmark coverage beyond the AA Intelligence Index
・StepFun's track record on open-weight delivery is unproven at this scale

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Kimi K3	1489	N/A	N/A	N/A
Qwen3.8-Max	1480	N/A	N/A	$2.50/$7.50
DeepSeek V4.1-Flash	N/A	N/A	N/A	N/A

Overview

Step 5 Preview is StepFun's 600B-parameter sparse Mixture-of-Experts model, previewed on September 21, 2026, with 27B active parameters per token. It carries a 1M-token context window and native vision input, and is priced at $1/$2.70 per million tokens. On the Artificial Analysis Intelligence Index it scores 44 — the same as Moonshot's Kimi K3, despite being roughly a quarter of its 2.8T size. StepFun plans to release open weights on October 15, 2026, which would make Step 5 one of the largest openly licensed models from a Chinese lab and a direct test of the 'API first, open weights later' release strategy.

Benchmarks & Performance

Step 5 Preview's only public benchmark to date is the Artificial Analysis Intelligence Index, where it scores 44 — matching Kimi K3 (44) at about a quarter of the parameter count, and ahead of Qwen3.8-Max (40) and DeepSeek V4.1-Flash (40). No SWE-Bench, GPQA, or Terminal-Bench numbers have been published. Independent evaluation is pending the open-weight release.

Detailed Comparison

**Step 5 Preview vs Kimi K3** - Kimi K3 is 2.8T parameters; Step 5 is 600B yet matches its AA index 44. - Kimi K3 already has open weights; Step 5's ship Oct 15. **Step 5 Preview vs Qwen3.8-Max** - Qwen3.8-Max scores AA index 40 and is a closed flagship; Step 5 leads on index and plans open weights. **Step 5 Preview vs DeepSeek V4.1-Flash** - Both target efficiency; V4.1-Flash is MIT and already downloadable, Step 5 is larger and pending.

Community Feedback

StepFun's profile rose through 2026 with Step 1/Step 2 agent models; Step 5 is its first frontier-scale entry. Developers are watching whether the October 15 open-weight drop actually ships at 600B, since Chinese labs have slipped open-weight dates before. The $1/$2.70 API price drew attention as aggressive for the size.

Use Cases

**1. Long-context multimodal agents** The 1M-token window with native vision suits document-heavy and screenshot-driven agent loops. Use it for research synthesis and multi-step workflows where context rarely resets. **2. Cost-sensitive coding assistance** At $1/$2.70 per 1M tokens it undercuts closed flagships for high-volume code tasks. Good for CI assistants and refactoring where extreme reasoning is not required. **3. Private deployment (post open weights)** If the Oct 15 open-weight release lands, Step 5 becomes a candidate for regulated or air-gapped inference at 600B scale — a rare option outside Western labs.

Latest News

**Sep 21, 2026**: Step 5 Preview announced. 600B MoE, 27B active, 1M context, native vision, $1/$2.70 per 1M. Open weights planned Oct 15. Artificial Analysis Intelligence Index 44, matching Kimi K3 at a quarter of the size.

Sources

Analysis generated: 2026-09-22