Step 5 Preview
Step 5 Preview is a large language model from the Chinese AI company StepFun (阶跃星辰), announced on 20 September 2026 as the company's new flagship model for agentic work. It is a sparse mixture-of-experts (MoE) model with 600 billion total parameters and 27 billion active per token, a 1-million-token context window, and vision input. [1][2] StepFun pitches it at software engineering and professional knowledge work, "with particular strength in finance," and frames the release around cost efficiency under the slogan "Advancing the Pareto Frontier." [1][2] At launch the model was available through StepFun's products and API under the model ID step-5-preview; StepFun said the weights would be released as open weights on 15 October 2026. [1][3]
Artificial Analysis, an independent benchmarking firm, gave Step 5 Preview a score of 44 on its Artificial Analysis Intelligence Index and measured an average cost of $0.72 per index task. [7] Most other performance figures published for the model come from StepFun's own evaluations.
Overview
| Item | Detail |
|---|---|
| Developer | StepFun (Shanghai) [1] |
| Announced | 20 September 2026 [2] |
| Architecture | Sparse MoE Transformer [1]; 92 layers, per StepFun as reported by Tencent Tech [8] |
| Parameters | 600B total, 27B active per token [1] |
| Context window | 1M tokens; maximum output 64K tokens [3] |
| Input / output | Text, images and video in; text out [3] |
| Reasoning effort | low, medium, high [3] |
| API model ID | step-5-preview [3] |
| API price (global platform) | $1.00 per 1M input tokens (cache miss), $0.05 (cache hit), $2.70 per 1M output tokens [4] |
| API price (China platform) | 7 yuan per 1M input tokens (cache miss), 0.35 yuan (cache hit), 20 yuan per 1M output tokens [5] |
| Weights | Promised as open weights on 15 October 2026 [1][2] |
Background
StepFun was co-founded in 2023 by Jiang Daxin, whom the South China Morning Post described as a former chief scientist of Microsoft's Asia research institute. [19] Its earlier flagship, Step-3, was a 321-billion-parameter vision-language MoE model with 38 billion active parameters, described in a technical report posted in July 2025. [12][13] In 2026 StepFun released two smaller "Flash" MoE models aimed at agent workloads, Step 3.5 Flash and Step 3.7 Flash. [14][16] Chinese coverage of the launch described Step 5 Preview as the third generation of this efficiency-focused line, after Step 3.5 Flash and Step 3.7 Flash. [8][9]
| Model | Date | Total / active parameters | Context | Notes |
|---|---|---|---|---|
| Step-3 | July 2025 | 321B / 38B | 64K (65,536 tokens) | Vision-language MoE; Apache 2.0 weights [12][13] |
| Step 3.5 Flash | February 2026 | 196B / 11B | 256K | Apache 2.0 weights; interleaved 3:1 sliding-window/full attention and 3-token multi-token prediction (MTP-3) [14][15] |
| Step 3.7 Flash | May 2026 | 198B / about 11B | 256K | 196B language backbone plus 1.8B vision encoder; Apache 2.0 weights [16][17][18] |
| Step 5 Preview | September 2026 | 600B / 27B | 1M | Text, image and video input; open weights promised for 15 October 2026 [1][3] |
With 27B of 600B parameters active (about 4.5%), Step 5 Preview has a lower active share than Step 3.5 Flash (11B of 196B, about 5.6%), a point Xinzhiyuan highlighted in its launch coverage. [1][9][15]
Architecture and training
StepFun's public model page gives only the headline specifications (600B total, 27B active, 1M context, vision input). [1] The design details below come from launch-day reports in Chinese tech media that attribute them to StepFun. [8][9][11]
- Narrow and deep. Instead of widening the network, Step 5 Preview uses a "narrow but deep" design with 92 Transformer layers. StepFun's reasoning is that agent tasks produce long prefills (search results, code execution output, tool returns), and a deeper network gives more room for implicit multi-hop reasoning while that context is processed. [8][9] Tencent Tech reported that StepFun said it had optimized training stability for the deeper network but had not disclosed its normalization or residual-structure details or ablation experiments. [8] Xinzhiyuan reported that the optimizations targeted hidden states, RMSNorm and gradients. [9]
- Sparse attention. On top of MoE expert sparsity, the model uses what StepFun calls Sparse GQA, a form of sparse attention built on grouped-query attention that uses a sparse index to pick the relevant parts of the history for each attention computation. To make that pay off on GPUs, it applies block-wise token merging, which StepFun says cuts the cost of the indexer and top-k selection steps to one eighth, and it merges heavily overlapping top-k selections of neighboring tokens to reduce fragmented computation. [8][9]
- Agent-built training data. StepFun says agents take part in its data pipeline, doing task design, difficulty and diversity control, and orchestration. Human experts set principles and make key judgments, and deterministic operations are wrapped as tools. The company says it added anti-hacking audits because stronger models tended to exploit evaluation loopholes, and that the pipeline has produced a million-scale set of verifiable hard tasks covering scientific reasoning, software engineering, machine-learning R&D and professional work. [8][9]
- Long-horizon reinforcement learning. According to Tencent Tech, StepFun says it achieved bit-wise alignment between training and inference in strict on-policy reinforcement learning, using deterministic operators and tree reduction so that tensor-parallel degree does not change results, at an end-to-end overhead below 10%. Other techniques listed include load-aware scheduling, MTP-3 speculative decoding (multi-token prediction), FP8 MoE, FP8 KV quantization-aware training, cross-node expert parallelism and KV cache offloading. StepFun says these speed up long-horizon RL end to end by more than three times, and that a sample-ledger mechanism keeps sample loss in the training pipeline below 1%. [8] The model is also trained to manage its own context as it nears the token budget, with a pretrained critic for token-level credit assignment and separate turn-level penalties for format errors and reward hacking. [8]
These are StepFun's claims as relayed by the press. Tencent Tech named the deployment requirements of a 92-layer model, the real speedup from Sparse GQA, and whether the open-weight release reproduces the API model's results as the main open questions for the October release. [8]
Release and availability
StepFun announced Step 5 Preview on X and on a dedicated model page on 20 September 2026, saying the model was "available today through our products and API" and that it "will be released with open weights on October 15." [1][2] Artificial Analysis lists the model's release date as 18 September 2026, two days before the formal announcement. [7]
The API documentation lists a 1M-token context window, maximum output of 64K tokens, text, image and video input (up to 60 images per request; video by URL, Base64 or Files API reference), streaming, tool calling, JSON Mode and JSON Schema structured output, prompt caching, and a reasoning-effort setting of low, medium or high. [3] The documentation notes that tools, search and code execution are supplied by the application that integrates the model. [3]
Pricing on StepFun's two platforms at launch:
| Platform | Input (cache miss) | Input (cache hit) | Output |
|---|---|---|---|
| Global (platform.stepfun.ai), per 1M tokens | $1.00 | $0.05 | $2.70 |
| China (platform.stepfun.com), per 1M tokens | 7 yuan | 0.35 yuan | 20 yuan |
Output tokens include both the reasoning trace and the final answer. [4][5] Standard API rate limits follow account tiers based on cumulative top-up, from 5 concurrent requests and 10 requests per minute at the lowest tier to 10,000 concurrent requests and 200,000 requests per minute at the highest. [4] Step 5 Preview is also included in Step Plan, StepFun's credit-based monthly subscription for coding tools and agent platforms, whose global tiers range from $6.99 to $99 per month. [6]
Benchmark results reported by StepFun
StepFun's model page includes a benchmark table comparing Step 5 Preview at "High" reasoning effort with five other models, each run at "Max": GLM-5.3, Kimi K3, GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5. [1] The table is StepFun's presentation. Benchmarks marked † were built by StepFun internally. On the model page, StepFun says GDPval-AA v2.1 scores are taken from Artificial Analysis as of 20 September 2026, and that DeepSWE v1.1 was run with the SWE-agent harness at temperature 1.0 and top_p 0.95. [1] For "HLE w/ tools" (marked ‡), Step 5 Preview and GLM-5.3 were run on the text-only subset while the other models were run on the full dataset, so StepFun notes those numbers are not directly comparable. [1] "n/a" means the page gives no score.
| Benchmark | Step 5 Preview (High) | GLM-5.3 (Max) | Kimi K3 (Max) | GPT-6 Astra (Max) | Claude Fable 5.1 (Max) | Claude Opus 5 (Max) |
|---|---|---|---|---|---|---|
| Reasoning and knowledge | ||||||
| GPQA Diamond | 93.5% | 91.7% | 93.5% | 96.1% | 93.7% | 93.2% |
| HLE | 46.5% | 42.3% | 46.9% | 54.7% | 59.1% | 54.9% |
| AA-LCR v1.1 | 88.3% | 79.7% | 88.7% | 80.7% | 85.3% | 79.3% |
| CritPt | 20.9% | 19.1% | 23.4% | 31.7% | 29.7% | 29.1% |
| Coding | ||||||
| DeepSWE v1.1 | 67.7% | 66.9% | 67.5% | 74.1% | 67.4% | 74.0% |
| Terminal-Bench v2.1 | 85.0% | 83.9% | 85.0% | 88.4% | 91.4% | 89.1% |
| Terminal-Bench v4 | 33.3% | 41.9% | 12.6% | 57.9% | 55.8% | 52.3% |
| CyberGym | 84.7% | 84.5% | 80.0% | n/a | n/a | n/a |
| SciCode | 58.9% | 59.0% | 59.5% | 56.5% | 63.1% | 56.4% |
| RoadmapBench | 54.3% | 54.1% | 55.4% | n/a | n/a | 68.3% |
| ProgramBench (pass rate) | 80.5% | 72.0% | 77.8% | 85.4% | 82.7% | 82.3% |
| SWE-Marathon v1.1 (partial score) | 72.7% | 67.4% | 84.4% | 77.3% | 80.2% | 85.6% |
| MLS-Bench-Lite | 40.5% | 37.3% | 48.3% | n/a | 50.3% | 49.8% |
| SWE-Atlas-QnA | 63.6% | 59.6% | 59.5% | 60.9% | n/a | 66.0% |
| SWE-Atlas-Test-writing | 50.8% | 50.4% | 50.4% | 51.1% | n/a | 60.3% |
| StepCodeBench † | 49.0% | 40.2% | 43.9% | 61.0% | n/a | 63.9% |
| StepCode-Bench-Daily † | 64.9% | 69.1% | 57.7% | n/a | n/a | 77.6% |
| StepCode-Bench-General † | 65.0% | 62.0% | 65.2% | 64.3% | n/a | 68.3% |
| General agent | ||||||
| GDPval-AA v2.1 (Elo) | 1566 | 1645 | 1524 | 1542 | 1735 | 1708 |
| τ³-Banking | 42.5% | 50.3% | 46.0% | 41.4% | 47.2% | 42.1% |
| AutomationBench-AA | 51.0% | 62.2% | 58.3% | 68.5% | 59.4% | 56.6% |
| AutomationBench (public) | 44.0% | 48.2% | 46.7% | n/a | n/a | n/a |
| AA-Briefcase v1.1 (Elo) | 1433 | 1526 | 1511 | 1569 | 1678 | 1673 |
| Toolathlon-Verified | 74.1% | 73.0% | 76.5% | n/a | 77.8% | 80.6% |
| MCP-Atlas | 85.6% | 86.8% | 85.3% | n/a | n/a | 87.0% |
| PresentBench | 76.8% | 74.5% | 75.6% | n/a | n/a | 77.3% |
| OfficeQA Pro | 60.3% | 59.1% | 62.6% | 67.7% | n/a | 64.7% |
| SpeadSheet v2 (as spelled on the page) | 29.4% | 30.5% | 31.9% | 31.4% | n/a | 32.8% |
| JobBench | 59.0% | 61.4% | 54.3% | n/a | n/a | 65.7% |
| Apex-Agents | 37.8% | 38.1% | 41.0% | n/a | n/a | 41.8% |
| Draco | 83.3% | 82.3% | 78.5% | 76.8% | 87.7% | 87.6% |
| BrowseComp | 88.7% | n/a | 91.2% | 91.5% | n/a | 90.2% |
| HLE w/ tools ‡ | 59.4% | 62.5% | 56.0% | 57.2% | 65.0% | 63.6% |
| FinStepBench-LiveSearch † | 74.5% | 73.3% | 70.9% | 74.5% | n/a | 76.2% |
| FinStepBench-CorporateValuation † | 60.6% | 56.1% | 60.6% | 77.3% | n/a | 69.7% |
| FinStepBench-FinanceDR † | 55.8% | 53.3% | 48.9% | 45.0% | n/a | 59.1% |
| FrontierFinance | 66.4% | 64.1% | 62.6% | 55.0% | n/a | 69.7% |
| Computer use | ||||||
| Agents' Last Exam (ALE-CLI) | 29.5% | 28.6% | 27.6% | 33.3% | n/a | 28.6% |
| Multimodal and document | ||||||
| MMMU-Pro | 76.0% | n/a (text only) | 81.0% | 87.0% | n/a | 85.0% |
| GDP.pdf | 14.8% | 11.2% | 22.0% | 31.0% | 26.2% | 21.6% |
In StepFun's own table, Step 5 Preview posts the top score in one row, CyberGym (84.7%), where only the three Chinese models have scores. It trails the US closed models on most coding, reasoning and agent rows, and against the two Chinese open-weight models it is ahead on some rows and behind on others. [1] For example, it beats Kimi K3 and GLM-5.3 on ProgramBench, StepCodeBench and FrontierFinance, trails GLM-5.3 on GDPval-AA, AutomationBench-AA and Terminal-Bench v4, and trails Kimi K3 on SWE-Marathon and MLS-Bench-Lite. [1] Several rows match the scores Artificial Analysis publishes for the model on its own evaluations (see below).
Coding and long-horizon demonstrations
StepFun describes Step 5 Preview's coding as broad rather than narrowly focused on code repair. It says that in evaluations by internal and external experts, about 70% of participants judged the model capable of autonomously solving coding tasks of moderately high complexity. [1]
StepCodeBench. StepFun built an internal benchmark, StepCodeBench, covering 553 independent repositories, 9 task categories, 20 application domains and 33 programming languages. Step 5 Preview scores 49.0% (avg@4). StepFun says it does best on bug repair, feature modification and refactoring, approaches Claude Opus 5 on low- and medium-difficulty tasks, and that "a meaningful gap to the frontier remains" on the hardest long-running tasks. [1]
WebDev, 3D and hardware. The model page shows the model building web interfaces and data visualizations, using Blender to create and refine 3D assets, and bringing those assets into Three.js web apps and games. Showcase cases include a room planner built from a bedroom photo, a 3D Flappy Bird, and an interactive atlas of the Kowloon-Canton Railway built from a 1922 travel guide. [1] Given documentation and user authorization, StepFun says the model can work with programmable hardware using cameras, COM ports, screenshots and simulated mouse input; the Chinese version of the page says the device demo ran continuously for more than three hours. [1]
GPU kernel optimization. StepFun gave the model 24 hours to optimize a multi-head latent attention (MLA) GPU kernel from scratch on one NVIDIA H100, with head dimension 512, batch size 1, 64 heads and 8,192 tokens. Each model got four independent attempts and the best run was reported. Step 5 Preview reached a peak of 508 TFLOPS for forward and backward after about 22 hours, against 493 TFLOPS for Claude Opus 5. The chart also includes Kimi K3 and GLM-5.3. [1]
Automated post-training. In a second 24-hour experiment, the model tried to improve a Qwen3-30B-A3B base model's AIME24 score through automated post-training, using an API annotator with access to production data. The resulting model reached 60% on the official AIME 2024 test, up from 53.3% before post-training. StepFun says this matched Claude Opus 5 while using fewer annotator tokens. [1]
Pokémon Red. As a long-horizon test outside coding, StepFun ran the model on Pokémon Red without any game-specific optimization. By the time of the announcement it had played more than 3,000 turns and about 6 million tokens of interaction; by turn 3,082 it had unlocked Cut, earned three Gym Badges and defeated Lt. Surge, which StepFun puts at roughly one-third of the main story. [1]
Professional work and finance
StepFun presents three examples of professional knowledge work. In a climate study covering 1,000 locations over 25 years, the model coordinated 950 web fetches in a single agent action and assembled 300,000 monthly records across 11 variables. In a diesel surcharge review it produced a 17-sheet analytical workbook with source data, cross-series reconciliation, regional panels, formulas and trend models. In a third task it produced an interactive research report. [1]
Finance is the area StepFun singles out. It built three internal benchmarks under the name FinStepBench: LiveSearch (retrieving and verifying timely financial information), CorporateValuation (turning financial data and assumptions into consistent forecasts and reproducible valuations) and DeepResearch (end-to-end research reports). It also reports results on FrontierFinance, an external benchmark with 220 expert-written questions and 11,543 evaluation criteria across six investment use cases. [1] The finance chart on the model page adds DeepSeek V4.1 Flash to the comparison:
| Benchmark | Step 5 Preview (High) | GLM-5.3 (Max) | Kimi K3 (Max) | DeepSeek V4.1 Flash | Claude Opus 5 (Max) | GPT-6 Astra (Max) |
|---|---|---|---|---|---|---|
| FrontierFinance | 66.4 | 64.1 | 62.6 | 63.0 | 69.7 | 55.0 |
| FinStepBench-LiveSearch | 74.5 | 73.3 | 70.9 | 76.7 | 76.2 | 74.5 |
| FinStepBench-CorporateValuation | 60.6 | 56.1 | 60.6 | 57.6 | 69.7 | 77.3 |
| FinStepBench-DeepResearch | 55.8 | 53.3 | 48.9 | 50.2 | 59.1 | 45.0 |
Source: StepFun model page. [1]
Independent evaluation
Artificial Analysis ran Step 5 Preview through its own test suite using StepFun's API. As of 23 September 2026 its model page showed: [7]
- Intelligence Index: 44 (43.7 before rounding) on version 4.3.2 of the index, ranked 32nd of 212 models in its comparison class, against a median of 25 for comparable models. The index combines ten evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1.
- Cost: $0.72 per Intelligence Index task on average.
- Verbosity: 160 million output tokens to complete the index, which Artificial Analysis calls "very verbose" against a median of 88 million.
- Speed: 83.1 output tokens per second and 3.62 seconds to first token on StepFun's API.
- Classification: listed as a proprietary model, with 600B total and 27B active parameters; the weights had not been released when the listing was accessed.
Artificial Analysis's per-evaluation scores match several rows in StepFun's table: Humanity's Last Exam 46.5%, AA-LCR 88.3%, CritPt 20.9%, Terminal-Bench 4.0 33.3%, SciCode 58.9%, GDPval-AA 1566, AutomationBench-AA 51.0% and GDP.pdf 14.8%. This strongly suggests those rows were taken from Artificial Analysis's measurements rather than run by StepFun, although StepFun's footnote attributes only the GDPval-AA scores to Artificial Analysis. StepFun's AA-Briefcase v1.1 figure (1433) differs slightly from the 1432 Artificial Analysis showed on 23 September 2026. [1][7] For MMMU-Pro, Artificial Analysis records 76.4% against StepFun's 76.0%. [1][7] Its AA-Omniscience results for the model are an index of 16.4, 41.5% accuracy and a 43.0% hallucination rate. [7]
StepFun's headline chart plots the model on Artificial Analysis's intelligence-versus-cost axes and labels it about 65% cheaper per task than GLM-5.3 (Max) and Kimi K3 (Max). [1][2] Tencent Tech reported a further StepFun claim that the model's per-task cost is one eighth that of Claude Opus 5, but, citing Artificial Analysis figures at the time of its report (about $0.71 per task for Step 5 Preview against about $2.03 for Claude Opus 5 at its highest reasoning setting), said the claim should be read as limited to the tasks and configurations StepFun chose. [8] In Artificial Analysis's data as accessed on 23 September 2026, the per-task costs are consistent with both StepFun claims: $0.72 for Step 5 Preview against about $2.00 for Kimi K3 (max) and GLM-5.3 (max) and $5.86 for Claude Opus 5 (max), as the table below shows. [7]
| Model (Artificial Analysis configuration) | Intelligence Index | Cost per index task | Open weights (per Artificial Analysis) |
|---|---|---|---|
| Step 5 Preview | 43.7 | $0.72 | No (promised) |
| Kimi K3 (max) | 43.6 | $2.00 | Yes |
| GLM-5.3 (max) | 44.8 | $2.01 | Yes |
| Xiaomi MiMo-V2.6-Pro | 46.3 | $0.13 | Yes |
| DeepSeek V4.1 Flash (max) | 39.5 | $0.27 | Yes |
| Claude Opus 5 (max) | 50.8 | $5.86 | No |
| GPT-6 Astra (max) | 52.7 | $3.26 | No |
| Claude Fable 5.1 (max) | 53.4 | $7.63 | No |
Source: Artificial Analysis model data, accessed 23 September 2026. [7]
StepFun described the model as ranking in the global top three open-source models on the Artificial Analysis index, a framing repeated by Xinzhiyuan; Silicon Star put it in the top two. [8][9][10] In Artificial Analysis's data as accessed on 23 September 2026, two open-weight models scored higher: GLM-5.3 (max) and Xiaomi's MiMo-V2.6-Pro, with Kimi K3 (max) just below. MiMo-V2.6-Pro, which Artificial Analysis lists as released on 21 September 2026 (one day after Step 5 Preview's announcement), scored higher than Step 5 Preview at a lower cost per task. [7] Tencent Tech noted that the model's heavy token use can partly offset its low per-token price. [8]
Reception
Coverage in the first days after launch came mainly from Chinese technology media; English-language coverage included Pandaily. Tencent Tech (腾讯科技) reported the specifications and training claims, noting what StepFun had not disclosed. [8] Pandaily summarized the architecture and the October open-weights plan. [11]
Two Chinese outlets ran hands-on tests. Xinzhiyuan (新智元), which had early access, described a working single-file Three.js flight game, a review of 12 bilingual contracts against 10 legal criteria, and a financial due-diligence memo with a formula-linked workbook, and reported that the model caught all 26 violations in the contract review and reproduced the data room's key figures exactly, with a 9-sheet workbook of 1,030 live formulas. [9] Silicon Star (硅星人) had the model build a WebGL island game, a Chrome summarization extension (whose 13 bundled unit tests passed) and an NVIDIA investment analysis with Excel models, a memo and an evidence list. It found that the model kept sources for key figures and listed data conflicts and assumptions, but that manual checks still turned up individual formula errors. Silicon Star concluded that the final numbers could not skip review by a professional. [10]
In the first days after launch, the main independent measurements were Artificial Analysis's; the other benchmark figures in this article come from StepFun. The open-weight release promised for 15 October 2026 would allow outside groups to run their own tests. [1][8]
See also
- StepFun
- Step-3
- Jiang Daxin
- Kimi K3
- GLM-5.3
- Xiaomi MiMo-V2.6
- DeepSeek V4
- MiniMax
- Artificial Analysis
- Mixture of Experts
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30StepFun, "Step 5 Preview: Advancing the Pareto Frontier" (model page; Chinese title "Step 5 Preview：向前一步，智能效率的新一代'帕累托前沿'"), 20 September 2026. stepfun.com/step-5-preview
- ^1 ^2 ^3 ^4 ^5 ^6StepFun (@StepFun_ai), "Introducing Step 5 Preview: Advancing the Pareto Frontier," X, 20 September 2026. x.com/...2101510462685003786
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8StepFun Documentation, "Step 5 Preview," StepFun Open Platform, accessed 23 September 2026. platform.stepfun.ai/...step-5-preview
- ^1 ^2 ^3StepFun Documentation, "Pricing and Rate Limits," StepFun Open Platform, accessed 23 September 2026. platform.stepfun.ai/...details
- ^1 ^2阶跃星辰开放平台文档, "定价与限速" (Pricing and rate limits, China platform), accessed 23 September 2026. platform.stepfun.com/...details
- ^StepFun Documentation, "Step Plan Overview," StepFun Open Platform, accessed 23 September 2026. platform.stepfun.ai/...overview
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Artificial Analysis, "Step 5 Preview - Intelligence, Performance & Price Analysis," accessed 23 September 2026. artificialanalysis.ai/...step-5
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15腾讯科技 (Tencent Tech), "阶跃发布Step 5 Preview：600B MoE瞄准长程Agent，10月将开源" (StepFun releases Step 5 Preview: 600B MoE aimed at long-horizon agents, to be open-sourced in October), Tencent News, 20 September 2026. news.qq.com/...20260920A06EUP00
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9新智元 (Xinzhiyuan), "刚刚，阶跃Step 5 Preview发布！一举杀进全球开源前三" (StepFun Step 5 Preview released, enters global open-source top three), via NetEase, 20 September 2026. 163.com/...L79ARRRU0511ABV6
- ^1 ^2硅星人 (Silicon Star), "阶跃发布 Step 5 Preview，大模型开始寻找新的平衡点" (StepFun releases Step 5 Preview as large models look for a new balance point), via NetEase, 21 September 2026. 163.com/...L7C734UE0511N33R
- ^1 ^2Pandaily, "StepFun Launches Step 5 Preview: 600B Sparse MoE, 1M Context, Weights Open Oct 15," 20 September 2026. pandaily.com/...step-5-preview-600b-moe-1m-context
- ^1 ^2StepFun, "Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding," arXiv:2507.19427, July 2025. arxiv.org/...2507.19427
- ^1 ^2StepFun, "step3" model card, Hugging Face. huggingface.co/...step3
- ^1 ^2StepFun, "Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters," arXiv:2602.10604, February 2026. arxiv.org/...2602.10604
- ^1 ^2StepFun, "Step-3.5-Flash" model card, Hugging Face. huggingface.co/...Step-3.5-Flash
- ^1 ^2StepFun, "Step 3.7 Flash: A high-efficiency Flash model for Real-World agents," StepFun blog, 29 May 2026. static.stepfun.com/...step-3.7-flash
- ^StepFun, "Step-3.7-Flash" model card, Hugging Face. huggingface.co/...Step-3.7-Flash
- ^StepFun Documentation, "Step 3.7 Flash," StepFun Open Platform, accessed 23 September 2026. platform.stepfun.ai/...step-3.7-flash
- ^Wency Chen, "Shanghai AI start-up founded by ex-Microsoft engineers bets on 'scaling law' to boost AI capabilities," South China Morning Post, June 2024. scmp.com/...bets-scaling-law-boost-ai-capabilities
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
1 revision · v2 · 4,138 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: xg07 independent adversarial verification 2026-09-23 (V2); 1 material + 4 minor fixed
Cite this page: AI Wiki. "Step 5 Preview." aiwiki.ai, updated 23 Sept 2026, fact-checked 23 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/step_5_preview