Step 5 Preview（阶跃星辰 StepFun） Step 5 Preview (StepFun)

来源Sources

为什么推荐Why we recommend it 收录理由：9/20 当日发布的中国厂商前沿旗舰，官方页信息完整（架构、上下文、基准表、可用性与开源排期），HN 与第三方榜同日有据可查，是今晚 Models 看板的头条卡。核实边界：参数量、上下文、基准分、开源日期一律为官方自述（取自 stepfun.com/step-5-preview 页面原文），本库未复跑任何基准；Intelligence Index 44、tok/s、$1 / $2.70 定价与 95% 缓存折扣来自 Artificial Analysis 第三方榜（官方公告页未列价，定价以阶跃 API 平台当日页为准）；OrcaRouter 博文是第三方产品博客，只作「权重仓为占位」这一事实的旁证，不为其路由产品背书。本库未实测 API 与 Studio。整理质量 4 星（信息齐、边界清），不代表基准分。 Why: a Chinese lab’s frontier flagship launched today (9/20) with a complete official page (architecture, context, benchmark table, availability, open-weights schedule), plus same-day HN and third-party-board coverage — tonight’s headline card on the Models board. Boundary: parameter counts, context window, benchmark scores, and the open-weights date are all vendor claims quoted from stepfun.com/step-5-preview; no benchmark was re-run here. Intelligence Index 44, tok/s, the $1 / $2.70 pricing, and the 95% cache discount come from the Artificial Analysis third-party board (the announcement page lists no prices; treat pricing per StepFun’s API platform page on the day). The OrcaRouter post is a third-party product blog, cited only as corroboration that the weights repo is a placeholder — no endorsement of its routing product. API and Studio were not tested here.

要点（官方称 / 社区称）Highlights (vendor / community claims)

• 官方称：面向 agentic 工作的新旗舰，稀疏 Mixture-of-Experts，600B 总参数、每 token 27B 活跃；支持 1M token 上下文与视觉输入；在软件工程、Agent 任务、专业知识工作与金融上「表现强劲」。 Vendor claim: a new flagship for agentic work built on a sparse Mixture-of-Experts — 600B total parameters with 27B active per token; 1M-token context and vision input; “performs strongly” across software engineering, agentic tasks, professional knowledge work, and finance.
• 官方称：官方页列出对照 GPT-6 Astra / Claude Opus 5 / Claude Fable 5.1 / GLM-5.3 / Kimi K3 的基准表（DeepSWE v1.1 67.7、Terminal-Bench v4 33.3、GPQA Diamond 93.5%、FrontierFinance 66.4 等），多数项目排在 GPT-6 Astra 与 Claude Opus 5 之后；带 † 的 StepCodeBench / FinStepBench 为其内部基准。本库未复跑任何一项。 Vendor claim: the official page tabulates results against GPT-6 Astra / Claude Opus 5 / Claude Fable 5.1 / GLM-5.3 / Kimi K3 (e.g. DeepSWE v1.1 67.7, Terminal-Bench v4 33.3, GPQA Diamond 93.5%, FrontierFinance 66.4), landing behind GPT-6 Astra and Claude Opus 5 on most rows; the †-marked StepCodeBench / FinStepBench are StepFun’s in-house benchmarks. None re-run here.
• 第三方榜（Artificial Analysis，非官方、非本库）：Intelligence Index 44（AA 页面当晚显示排名 #24 / 200），输出约 99.8 tok/s，标价 $1.00 / 百万输入 token、$2.70 / 百万输出 token、缓存折扣 95%；AA 把该模型日期记为 2026-09-18，比官方公告早两天。定价以阶跃星辰 API 平台当日页为准，官方公告页本身未列价。 Third-party board (Artificial Analysis — neither vendor nor this library): Intelligence Index 44 (ranked #24 / 200 on the AA page tonight), ~99.8 output tok/s, listed at $1.00 per 1M input tokens / $2.70 per 1M output tokens with a 95% cache discount; AA dates the model 2026-09-18, two days before the vendor announcement. Treat pricing per StepFun’s API platform page on the day — the announcement page itself lists no prices.
• 官方称：「Step 5 Preview 今天起可通过我们的产品与 API 使用，模型将于 10 月 15 日以开源权重发布」。截至今晚，Hugging Face 的 stepfun-ai/Step-5-Preview-BF16 仓库对本库返回 401；OrcaRouter 博文称该仓库只有一个 .gitattributes、无权重 / 许可 / 模型卡。开源日期是厂商排期，不是已发生的事实。 Vendor claim: “Step 5 Preview is available today through our products and API. The model will be released with open weights on October 15.” As of tonight the Hugging Face repo stepfun-ai/Step-5-Preview-BF16 returned 401 to this library; the OrcaRouter post says it holds a single .gitattributes with no weights, license, or model card. The open-weights date is a vendor schedule, not a shipped artifact.
• 官方页 demo 自述（未核实）：Room Planner 从卧室照片生成可编辑 3D 房间；Blender + Three.js 做 3D 资产与网页游戏；24 小时自动后训练把 Qwen3-30B-A3B 的 AIME24 从 53.3% 提到 60%；无专门优化下玩 Pokémon 超 3,000 回合、6M token。均为厂商展示，本库未复现。 Official-page demos, as described by the vendor (unverified): a Room Planner that turns a bedroom photo into an editable 3D room; Blender + Three.js for 3D assets and browser games; a 24-hour automated post-training run lifting Qwen3-30B-A3B’s AIME24 from 53.3% to 60%; a Pokémon run past 3,000 turns / 6M tokens without game-specific tuning. All vendor showcases, not reproduced here.