StepFun Step 5 Preview: 600B Sparse MoE, 27B Active Parameters, and the Pareto Frontier for Agentic Coding
On September 18-20, 2026, StepFun released Step 5 Preview β a 600B-parameter sparse MoE model with only 27B activated per token, achieving 67.7% on DeepSWE v1.1, 33.3% on Terminal-Bench 4.0, and 99.8 tokens/sec output speed at $1/M input pricing, with open weights scheduled for October 15, 2026.
StepFun Step 5 Preview: 600B Sparse MoE, 27B Active Parameters, and the Pareto Frontier for Agentic Coding
Executive Summary
On September 18-20, 2026, Beijing-based AI lab StepFun (ιΆθ·ζθΎ°) released Step 5 Preview, a sparse Mixture-of-Experts (MoE) model that represents one of the most aggressive efficiency plays in the frontier AI landscape. With 600 billion total parameters but only 27 billion activated per token (~4.5% sparsity), Step 5 Preview achieves near-frontier agentic coding performance at a fraction of the inference cost of comparable models.
The release delivers across three dimensions: architectural efficiency, benchmark competitiveness, and pricing disruption. Architecturally, Step 5 Preview uses a 92-layer "narrow but deep" Transformer with Sparse Grouped-Query Attention (GQA) and block-wise token merging to support a 1M-token context window while reducing indexer and top-k selection costs to approximately one-eighth of a denser baseline. On benchmarks, it achieves 67.7% on DeepSWE v1.1 (leading all open-weight models), 33.3% on Terminal-Bench 4.0 (2.6Γ higher than Kimi K3), and 80.5% on ProgramBench, while scoring 44 on the Artificial Analysis Intelligence Index β on par with Kimi K3 Max (2.8T parameters) and competitive with Qwen3.8-Max.
Pricing is set at Β₯7/M input (~$1/M) and Β₯20/M output (~$2.70/M) with Β₯0.35/M cached input (~$0.05/M) β roughly 10Γ cheaper on input and 18.5Γ cheaper on output than the US frontier models (GPT-6 Astra at $10/$50, Claude Fable 5.1 at $10/$50). API access is available immediately, with open weights scheduled for October 15, 2026 under the StepFun Community License.
This article extends the September 2026 frontier model survey alongside Qwen3 8 Max First Open Source Flagship Hybrid Deltanet Moe Coding Cowork 2026 09 18, Claude Fable 5 1 Mythos 5 1 Scientific Frontier Agentic Coding Enterprise Frontier Safeguards 2026 09 17, Gpt 6 Astra Critical Cyber Capabilities Computer Use Frontier Alignment Monitorability 2026 09 16, and Gemini 3 8 Flash Agentic Coding Cyber Security Fairwind Program 2026 09 15.
1. The StepFun Story: From Step-1 to Step 5 in 3 Years
1.1 Company Background
StepFun (ιΆθ·ζθΎ°) was founded in April 2023 by Jiang Daxin, formerly Google Brain's director of applied AI research and creator of the Inception network. The company's mission is "ζΊθ½ιΆθ·οΌεεζ―δΈͺδΊΊηε―θ½" (Intelligent leap, ten times everyone's potential).
StepFun's model evolution has been aggressive:
| Model | Release | Architecture | Key Feature |
|---|---|---|---|
| Step-1 | 2023 | Dense | 100B+ parameter bilingual (Chinese/English) model |
| Step-1V | 2023 | Dense | First multimodal variant |
| Step-3 | July 2025 | MoE (321B/38B) | First MoE, vision-language reasoning |
| Step-3.5-Flash | Feb 2026 | MoE (196B/11B) | Open-source, training code released (SteptronOss) |
| Step-3.7-Flash | May 2026 | MoE | Real-world agentic optimization |
| Step-5 Preview | Sep 2026 | MoE (600B/27B) | Flagship, 1M context, Pareto frontier |
Notably, StepFun skipped the entire Step-4.x line, jumping directly from Step-3.7-Flash to Step-5 β a decision reflecting the magnitude of improvement achieved in this release.
1.2 The "Pareto Frontier" Philosophy
StepFun's core philosophy for Step 5 is the "Pareto Frontier" β achieving the optimal balance between intelligence and cost. While previous scaling efforts focused on trading more compute for stronger intelligence, StepFun argues that the next phase requires improving the efficiency of converting compute into intelligence.
2. Architecture: 600B Total, 27B Active, 92 Layers Deep
2.1 Sparse MoE at Scale
Step 5 Preview's architecture is designed around a radical separation of total model capacity from per-token computation:
| Component | Specification |
|---|---|
| Total Parameters | 600 billion |
| Active Parameters | 27 billion per token |
| Sparsity | ~4.5% (95.5% inactive per token) |
| Layers | 92 (narrow-deep Transformer) |
| Attention | Sparse GQA with block-wise token merging |
| Context Window | 1,000,000 tokens |
| Input Modalities | Text, Image, Video |
| Output Modalities | Text |
The 27B activation figure does not mean Step 5 Preview behaves like a conventional dense 27B model. The larger 600B parameter pool gives the routing system access to many more specialized experts across different tokens and tasks. Different expert subsets are selected as the model processes different inputs, providing far richer representational capacity than a fixed 27B model.
2.2 Narrow-Deep Design
Step 5 Preview uses a 92-layer "narrow but deep" configuration. This design creates longer information propagation paths for implicit multi-hop reasoning during long prefill operations. In contrast to wider, shallower models that process information through more parallel pathways, the narrow-deep design forces information through more sequential transformation layers, potentially enabling more sophisticated reasoning chains.
2.3 Sparse GQA with Block-Wise Token Merging
To handle the 1M-token context window efficiently, Step 5 Preview introduces Sparse GQA with block-wise token merging:
- Sparse indexing selects only historical information relevant to the current task
- Block-wise token merging groups similar tokens into representative blocks before attention computation
- This mechanism reduces the number of tokens that actually enter attention computation
- StepFun states this cuts indexer and top-k selection costs to approximately one-eighth of a denser baseline
2.4 Multimodal Encoder
The model incorporates a unified multimodal encoder that processes text, images, and video frames into a shared latent space:
- Images: JPG/JPEG, PNG, WebP, static GIF; up to 60 images per request;
lowandhighdetail levels - Video: MP4, QuickTime, Matroska; β€128MB per file; β€5 minutes recommended
- Video processing: Adaptive frame rate sampling with temporal attention for motion understanding
3. Benchmark Performance: Leading Open-Weight Models
3.1 Coding & Software Engineering
Step 5 Preview leads all open-weight models on key coding benchmarks:
| Benchmark | Step 5 Preview | Kimi K3 Max | GLM-5.3 Max | Claude Opus 5 | GPT-6 Astra |
|---|---|---|---|---|---|
| DeepSWE v1.1 | 67.7% | 67.5% | 66.9% | 74.0% | 74.1% |
| StepCodeBench | 49.0% | 43.9% | 40.2% | 63.9% | 61.0% |
| ProgramBench | 80.5% | 77.8% | 72.0% | 82.3% | 85.4% |
| Terminal-Bench 4.0 | 33.3% | 12.6% | 41.9% | 52.3% | 57.9% |
| Agents' Last Exam | 29.5% | 27.6% | 28.6% | 28.6% | 33.3% |
On DeepSWE v1.1, Step 5 Preview leads all open-weight models (67.7% vs. Kimi K3's 67.5%), trailing only the larger closed-source models. On Terminal-Bench 4.0, it achieves 33.3% β 2.6Γ higher than Kimi K3 (12.6%) and 1.24Γ higher than DeepSeek V4.1-Flash (26.8%).
3.2 Financial & Professional Knowledge
| Benchmark | Step 5 Preview | Kimi K3 Max | GLM-5.3 Max | Claude Opus 5 | GPT-6 Astra |
|---|---|---|---|---|---|
| FrontierFinance | 66.4% | 62.6% | 64.1% | 69.7% | 55.0% |
| DRACO | 83.3% | 78.5% | 82.3% | 87.6% | 76.8% |
| GDPval-AA v2 | 1571 | 1548 | 1634 | 1735 | 1580 |
On FrontierFinance, Step 5 Preview outperforms both Kimi K3 and GLM-5.3 by a significant margin and beats GPT-6 Astra (66.4% vs. 55.0%), suggesting particular strength in financial analysis and reasoning tasks.
3.3 Performance Characteristics
| Metric | Step 5 Preview | GLM-5.3 Max | Claude Opus 5 |
|---|---|---|---|
| Output Speed | 99.8 tokens/sec | 72.1 tokens/sec | β |
| Time to First Token | 2.96 seconds | 2.99 seconds | 56.84 seconds |
| Intelligence Index | 44 | β | β |
The 2.96-second time to first token is dramatically faster than Claude Opus 5's 56.84 seconds at max effort, making Step 5 Preview particularly suitable for interactive agentic workflows where low latency matters.
4. Agentic Capabilities: 24-Hour Autonomous Execution
4.1 GPU Kernel Optimization
In a landmark demonstration of sustained agentic execution, Step 5 Preview was tasked with autonomously optimizing an H100 GPU kernel for up to 24 consecutive hours:
- Task: Independently modify code, run tests, compare results, iterate based on performance
- Duration: ~22 hours of continuous execution
- Result: Reached 508 TFLOPS
- Comparison: Claude Opus 5 achieved 493 TFLOPS in the same experiment
This demonstrates Step 5 Preview's ability to sustain productive work over extended periods without human intervention β a critical capability for autonomous coding agents.
4.2 Automated Post-Training
In another 24-hour experiment, Step 5 Preview autonomously improved the accuracy of Qwen3-30B-A3B on AIME24 from 53.3% to 60% through automated post-training experiments. This showcases the model's capacity for self-directed research and optimization β effectively acting as an AI research assistant that can improve other models.
4.3 Real-World Demonstrations
StepFun demonstrated the model across several complex projects:
| Domain | Task | Duration | Outcome |
|---|---|---|---|
| Hardware | ESP32 development board modifications | 3+ hours | Full hardware programming cycle |
| Frontend | Full-stack design with 3D asset generation | Multi-hour | Visual design + code |
| Finance | End-to-end investment research | Multi-hour | Data gathering β report |
| Research | Autonomous research assistant | Extended | Paper reading β experiments β summary |
5. Pricing Analysis: The Cost Disruption
5.1 API Pricing
| Metric | Step 5 Preview | Qwen3.8-Max | Claude Fable 5.1 | GPT-6 Astra | Gemini 3.8 Flash |
|---|---|---|---|---|---|
| Input | Β₯7/M (~$1.00) | Β₯14/M (~$2.00) | $10.00 | $10.00 | $0.75 |
| Output | Β₯20/M (~$2.70) | Β₯42/M (~$6.00) | $50.00 | $50.00 | $3.75 |
| Cache Read | Β₯0.35/M (~$0.05) | Β₯1.75/M (~$0.25) | $0.25 | $1.00 | β |
| Context | 1M | 1M | 200K | 1.05M | 1M |
Step 5 Preview is 10Γ cheaper on input and 18.5Γ cheaper on output than the US frontier models (Fable 5.1, GPT-6 Astra), while matching their context windows. The cache price of ~$0.05/M is 5Γ cheaper than Anthropic's $0.25/M rate.
5.2 Cost Comparison on Agentic Workloads
For a typical agentic coding task with 500K input tokens (80% cache hit) and 50K output tokens:
| Model | Estimated Cost |
|---|---|
| Step 5 Preview | ~$0.64 |
| Qwen3.8-Max | ~$1.95 |
| Gemini 3.8 Flash | ~$413 |
| Claude Fable 5.1 | ~$2,725 |
| GPT-6 Astra | ~$5,500 |
The pricing difference is dramatic. For high-volume agentic workloads where output tokens dominate, Step 5 Preview offers orders of magnitude cost savings compared to US frontier models.
5.3 Rate Limits
StepFun implements tiered rate limits based on accountε εΌ (top-up) amount:
| Tier | Top-up | Concurrency | RPM | TPM |
|---|---|---|---|---|
| V0 | Β₯0 | 5 | 10 | 5M |
| V1 | Β₯100 | 100 | 1,000 | 20M |
| V2 | Β₯500 | 200 | 5,000 | 30M |
| V3 | Β₯2,000 | 400 | 10,000 | 40M |
| V4 | Β₯5,000 | 1,000 | 20,000 | 50M |
| V5 | Β₯10,000 | 10,000 | 200,000 | 100M |
Free accounts (V0) are limited to 5 concurrent requests and 10 RPM, but paying Β₯100 (~$14) unlocks 100 concurrency and 1,000 RPM.
6. Deployment & Implementation
6.1 API Access
Step 5 Preview is available via OpenAI-compatible API:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_STEP_API_KEY",
base_url="https://api.stepfun.com/v1",
)
response = client.chat.completions.create(
model="step-5-preview",
messages=[{"role": "user", "content": "Optimize this GPU kernel for H100..."}],
reasoning_effort="high", # low / medium / high / xhigh
max_tokens=2048,
)
6.2 Self-Hosted Deployment (Open Weights Coming Oct 15)
The BF16 checkpoint is available on Hugging Face (TypeSafeAI/Step-5-Preview-BF16), with open weights under StepFun Community License scheduled for October 15, 2026:
# vLLM deployment
vllm serve TypeSafeAI/Step-5-Preview-BF16 \
--trust-remote-code \
--tensor-parallel-size 8 \
--max-model-len 1000000 \
--enable-reasoning \
--reasoning-parser stepfun
# SGLang deployment
python -m sglang.launch_server \
--model-path TypeSafeAI/Step-5-Preview-BF16 \
--trust-remote-code \
--tp 8 \
--context-length 1000000 \
--reasoning-parser stepfun
Hardware requirements (estimated for BF16):
- Minimum: 8Γ NVIDIA H100 (80GB) for 1M context
- Memory: ~1.2 TB weight data at BF16 (600B Γ 2 bytes)
- FP8: 4Γ H100 (80GB) expected (coming soon)
6.3 Reasoning Effort Control
Step 5 Preview supports four reasoning effort levels:
| Effort Level | Use Case | Estimated Latency |
|---|---|---|
| xhigh | Complex reasoning, research, optimization | Highest |
| high | Default for coding, professional tasks | High |
| medium | General tasks, balanced accuracy/speed | Moderate |
| low | Simple queries, rapid iteration | Lowest |
7. Integration with Prior Research
7.1 Positioning in the September 2026 Landscape
Step 5 Preview completes the picture of Chinese frontier models in the September 2026 release wave:
| Dimension | Qwen3.8-Max | Step 5 Preview | DeepSeek V4.1-Flash |
|---|---|---|---|
| Release Date | Aug 26 | Sep 18-20 | Sep 10 |
| Total Params | 2.4T | 600B | 552B |
| Active Params | 95B | 27B | 8B-16B |
| Open Weights | Yes (FP8) | Oct 15 (BF16) | Yes (MIT) |
| DeepSWE v1.1 | β | 67.7% | 74.2% |
| Terminal-Bench 4.0 | β | 33.3% | 26.8% |
| Pricing | $2/$6 | $1/$2.70 | $0.30/$1.20 |
| Context | 1M | 1M | 1M |
| Multimodal | No (flagship) | Yes | Yes |
See Qwen3 8 Max First Open Source Flagship Hybrid Deltanet Moe Coding Cowork 2026 09 18 for Qwen3.8-Max and Deepseek V4 1 Flash Causal Encoder Decoder Kv Cache Compression Agentic Coding 2026 09 14 for DeepSeek V4.1-Flash.
7.2 The Efficiency Frontier
Step 5 Preview occupies a unique position in the efficiency frontier:
| Model | Active Params | Intelligence Index | Cost per 1M Output |
|---|---|---|---|
| Step 5 Preview | 27B | 44 | $2.70 |
| Qwen3.8-Max | 95B | ~44+ | $6.00 |
| DeepSeek V4.1-Flash | 8B-16B | ~40+ | $1.20 |
| Claude Fable 5.1 | Undisclosed | ~66 | $50.00 |
| GPT-6 Astra | Undisclosed | ~67 | $50.00 |
Step 5 Preview achieves near-Qwen3.8-Max intelligence with 71% fewer active parameters and 55% lower output pricing, representing a significant improvement in the compute-to-intelligence conversion ratio.
7.3 Comparison with US Frontier Models
| Dimension | GPT-6 Astra | Claude Fable 5.1 | Step 5 Preview |
|---|---|---|---|
| DeepSWE v1.1 | 74.1% | 74.0% | 67.7% |
| Terminal-Bench 4.0 | 57.9% | 55.8% | 33.3% |
| Computer Use | 72.6% OSWorld | 77.9% OSWorld | Not tested |
| Cybersecurity | Critical threshold | Mythos 5.1 | Not emphasized |
| Pricing | $10/$50 | $10/$50 | $1/$2.70 |
| Context | 1.05M | 200K | 1M |
| Open Weights | No | No | Oct 15 |
US models dominate on absolute capability (cybersecurity, computer use, agentic science); Step 5 Preview dominates on cost-efficiency and offers open weights.
8. Key Takeaways
-
Efficiency-first scaling works. Step 5 Preview demonstrates that 600B total parameters with 27B active can compete with models 5-10Γ larger in active parameters, validating StepFun's Pareto Frontier philosophy.
-
Chinese frontier models are converging. With Qwen3.8-Max (2.4T/95B), Step 5 Preview (600B/27B), and DeepSeek V4.1-Flash (552B/8-16B), China now offers three distinct open-weight frontier options spanning different efficiency points.
-
Pricing disruption is real. At $1/M input and $2.70/M output, Step 5 Preview makes frontier-level agentic coding accessible at prices previously only available from mid-tier models.
-
24-hour autonomous execution is practical. The GPU kernel optimization demo (508 TFLOPS, beating Claude Opus 5) proves that sparse MoE models can sustain long-horizon autonomous work without degradation.
-
Open weights coming October 15. The BF16 checkpoint is already available, and the StepFun Community License release will enable self-hosting, fine-tuning, and independent evaluation.
-
Multimodal at frontier efficiency. Native text, image, and video support in a 27B-active model is a significant advantage for multimodal agentic workflows.
-
Financial reasoning strength. Outperforming GPT-6 Astra on FrontierFinance (66.4% vs. 55.0%) suggests Step 5 Preview may be particularly well-suited for financial analysis and professional knowledge work.
9. References & Resources
Official Sources
- StepFun: Step 5 Preview Announcement β Official product page
- StepFun Platform: Step 5 Preview Documentation β Model specifications and API guide
- StepFun Platform: Pricing & Rate Limits β Detailed pricing table
- Hugging Face: TypeSafeAI/Step-5-Preview-BF16 β Model card with full benchmark tables
- GitHub: StepFun AI β Official GitHub organization
Inference Frameworks
- vLLM Step 5 Preview β vLLM deployment guide
- SGLang Step 5 Preview β SGLang deployment guide
Related Research
- Qwen3 8 Max First Open Source Flagship Hybrid Deltanet Moe Coding Cowork 2026 09 18 β Qwen3.8-Max deep-dive
- Claude Fable 5 1 Mythos 5 1 Scientific Frontier Agentic Coding Enterprise Frontier Safeguards 2026 09 17 β Anthropic's September release
- Gpt 6 Astra Critical Cyber Capabilities Computer Use Frontier Alignment Monitorability 2026 09 16 β OpenAI's GPT-6 generation
- Gemini 3 8 Flash Agentic Coding Cyber Security Fairwind Program 2026 09 15 β Google's iterative Flash strategy
- Deepseek V4 1 Flash Causal Encoder Decoder Kv Cache Compression Agentic Coding 2026 09 14 β DeepSeek-V4.1-Flash deep-dive
External Benchmarks
10. Future Directions
10.1 Open Weights Release (October 15)
The October 15 open weights release is the next critical checkpoint:
- License terms: The StepFun Community License will determine commercial deployment permissions
- Quantization: FP8 and lower-precision variants will enable deployment on fewer GPUs
- Fine-tuning: Community fine-tunes for specific domains (legal, medical, financial)
- Independent evaluation: Third-party benchmarks beyond StepFun's own evaluations
10.2 Step 5 Final Release
The "Preview" designation suggests a final Step 5 release is coming:
- Architecture refinements: Based on early API usage feedback
- Additional benchmarks: More comprehensive evaluation across domains
- Safety improvements: Alignment and safety enhancements
- Multimodal output: Image and video generation capabilities
10.3 Competitive Response
Step 5 Preview's pricing may pressure other providers:
- Qwen: Potential price reduction for Qwen3.8-Max API
- DeepSeek: Possible V4.2 or V5 release responding to StepFun's efficiency gains
- US labs: May reconsider pricing strategies for mid-tier models
10.4 Watch For
- Open weights release (October 15, 2026) β License terms and deployment feasibility
- Step 5 final release β Expected Q4 2026 based on the Preview timeline
- Community benchmarks β Independent validation of the 24-hour autonomous execution claims
- Multimodal fine-tunes β Domain-specific adaptations for financial analysis, code review, and research
- Hardware optimization β FP8 deployment on 4Γ H100 and consumer GPU support
Citation:
@misc{step5preview,
title = {{Step 5 Preview}: Advancing the Pareto Frontier},
url = {https://www.stepfun.com/step-5-preview},
author = {{StepFun Team}},
month = {September},
year = {2026}
}
This article was researched using official StepFun sources: the StepFun product page, platform documentation, pricing page, Hugging Face model card, and GitHub repository. All benchmark figures are from StepFun's published evaluations and the Hugging Face model card.
π Referenced by
- π¬Grok 4.7: Long-Running Agents, Frontier Coding at $2/M, and xAI's New Safeguard Stack2026-09-25T00:00:00.000Z
- π¬GPT-6 Sol and Luna: 50% Price Cuts, Agentic Coding at Scale, and the New Cost-Efficiency Frontier2026-09-24T00:00:00.000Z
- π¬Claude Opus 5.5: Fable 5.1 Performance at 40% Less Cost, the First Release After the Pacing Call2026-09-23T00:00:00.000Z
- π September 21: StepFun Step 5 Preview & AI News Weekly β The Efficiency Frontier & This Week's Biggest Stories2026-09-21T00:00:00.000Z
- πWiki Index2026-06-17T00:00:00.000Z
- πWiki Log2026-06-17T00:00:00.000Z