Source量子位•Freshcollected in 3h
Step 5 Preview Brings a 27B Active Model
#open-weights#active-parameters#model-evaluationstep-5-previewstep 5 previewstep
A new open model claims strong performance with only 27B active parameters.
30-Second TL;DR
What Changed
Step 5 Preview is presented as an open-source model.
Why It Matters
A capable open model with 27B active parameters could offer a practical balance between quality and inference cost. More evidence is needed before deciding whether it is suitable for production workloads.
What To Do Next
Obtain the Step 5 Preview weights and run standardized generation, latency, and memory tests before replacing your current open model.
Who should care:Developers & AI Engineers
Key Points
- •Step 5 Preview is presented as an open-source model.
- •The reported active parameter count is 27 billion.
- •The available excerpt does not include benchmark scores or deployment requirements.
Key numbers$1.00$2.70
Deep Insight
Background and context from public sources — not the original article. 13 sources cited.
Enhanced Key Takeaways
- •Step 5 Preview is developed by Chinese AI startup StepFun (阶跃星辰) as a massive 600B total parameter sparse Mixture-of-Experts (MoE) model that skips the 4.x generation.
- •The model scored 44 points on the Artificial Analysis Intelligence Index, ranking among the top two open-source models globally and surpassing Moonshot's Kimi K3.
- •API inference is aggressively priced at $1.00 per 1M input tokens and $2.70 per 1M output tokens, cutting compute costs to approximately 1/8th of Anthropic's Claude Opus 5.
- •It natively supports multimodal inputs (text, image, and video) alongside a 1M-token context window designed for long-horizon agentic execution.
- •In autonomous engineering benchmarks, Step 5 Preview optimized an MLA GPU kernel over a 24-hour cycle to reach 508 TFLOPS throughput, outperforming Claude Opus 5 (493 TFLOPS).
Competitor Analysis
Step 5 Preview
- Developer
- StepFun
- Architecture / Active Params
- MoE (600B Total / 27B Active)
- Context Window
- 1M tokens
- Benchmark (Intelligence Index)
- 44 (Top 2 Open-Source)
- Pricing (Input/Output per 1M)
- $1.00 / $2.70
Kimi K3
- Developer
- Moonshot AI
- Architecture / Active Params
- Proprietary
- Context Window
- Extended
- Benchmark (Intelligence Index)
- Trailing Step 5 Preview
- Pricing (Input/Output per 1M)
- Undisclosed / Competitive
Claude Opus 5
- Developer
- Anthropic
- Architecture / Active Params
- Frontier Closed-Source
- Context Window
- Long-context
- Benchmark (Intelligence Index)
- Leading Tier (493 TFLOPS kernel baseline)
- Pricing (Input/Output per 1M)
- ~8x higher than Step 5
| Model | Developer | Architecture / Active Params | Context Window | Benchmark (Intelligence Index) | Pricing (Input/Output per 1M) |
|---|---|---|---|---|---|
| Step 5 Preview | StepFun | MoE (600B Total / 27B Active) | 1M tokens | 44 (Top 2 Open-Source) | $1.00 / $2.70 |
| Kimi K3 | Moonshot AI | Proprietary | Extended | Trailing Step 5 Preview | Undisclosed / Competitive |
| Claude Opus 5 | Anthropic | Frontier Closed-Source | Long-context | Leading Tier (493 TFLOPS kernel baseline) | ~8x higher than Step 5 |
Technical Deep Dive
- MoE Configuration: Total scale of 600 billion parameters with a sparse ~4.5% activation ratio, yielding 27 billion active parameters per token.
- Depth and Topology: Employs a 'narrow but deep' 92-layer Transformer backbone to maintain representational capacity at low per-token compute.
- Context and Modality: Integrated 1,000,000-token context support with native multimodal ingestion covering text, high-resolution imagery, and video.
- Agentic Endurance: Verified to sustain multi-hour tool orchestration via API, including autonomous 3D scene construction in Blender and multi-step GPU kernel optimization.
Future ImplicationsAI analysis grounded in cited sources
Extreme MoE sparsity will emerge as the standard blueprint for frontier open-source agent models.
Activating only 4.5% of total parameters allows Step 5 Preview to deliver 600B-grade reasoning and top-tier GPU kernel optimization at a fraction of standard inference overhead.
Price erosion will accelerate sharply in frontier-class multimodal API services.
Offering near-Opus level agentic capability at $1.00/$2.70 per million tokens forces competing frontier providers to slash API markups on complex reasoning tasks.
Timeline
2026-09
StepFun unveils Step 5 Preview with 600B total/27B active MoE architecture and 1M context window
2026-09
API access and preview platform launched; open weights release scheduled for October 2026
- 2026-09StepFun unveils Step 5 Preview with 600B total/27B active MoE architecture and 1M context window
- 2026-09API access and preview platform launched; open weights release scheduled for October 2026
Sources (13)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
1493179qbitai.com2Step 5 Preview Bf16huggingface.co37687630544473636910juejin.cn4ycombinator.comnews.ycombinator.com5Step 5 Preview Leakorcarouter.ai616688firecat-web.com77687630544473636910juejin.cn8Step 5 Previewstepfun.com9L7bmj33005568w0a163.com10Stepfun Step 5 Previewoschina.net11Stepfun Ships Step 5 Preview API a 600b Moe at 1270 That Scores 44 onaiweekly.co12163.comc.m.163.com13geekpark.netabout.geekpark.net
Weekly AI Recap
Read this week's curated digest of top AI events →
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 量子位 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.