Scale-up possibilities for everyone. API Platform: https://xitter.cf/t.co/XlaiCy4DGm HuggingFace: https://xitter.cf/t.co/isMgbv7XSm

Joined February 2025
sf, join us this thursday for step 5 preview launch night with stepfun we’ll dig into the model, building + serving agent workloads, and what’s worth building as models keep improving, with dan fu, hongyu zhou, mar hershenson, saoud rizwan, joey wang rsvp: luma.com/12t5e4lc
4
4
15
2,836
congrats to the StepFun team on releasing Step 5 Preview open weights today we’re excited to be a launch partner and to support another StepFun release as they bring Step 5 Preview to developers
Step 5 Preview is now open weights. BF16 and FP8 weights are live on Hugging Face and ModelScope, together with our technical blog. • Sparse MoE: 600B total, 27B active per token • 1M-token context window • Native text and vision input • Built for long-horizon agentic work: coding, software engineering, professional knowledge work and finance Frontier intelligence, open weights. Build what's next. Hugging Face: huggingface.co/stepfun-ai/St… ModelScope: modelscope.ai/models/stepfun… GitHub: github.com/stepfun-ai/Step-5…
5
2
25
3,077
StepFun retweeted
RadixArk
@radixarkOct 15
Congratulations to @StepFun_ai on Step 5 Preview going open weights! RadixArk is proud to help power StepFun's first-party production API with SGLang at scale. Production inference means great token economics under strict latency requirements. SGLang and its cache-aware routing stack now bring that to Step 5 Preview in proven production traffic. Building with Step 5 Preview? Deploy with SGLang today!
Step 5 Preview is now open weights. BF16 and FP8 weights are live on Hugging Face and ModelScope, together with our technical blog. • Sparse MoE: 600B total, 27B active per token • 1M-token context window • Native text and vision input • Built for long-horizon agentic work: coding, software engineering, professional knowledge work and finance Frontier intelligence, open weights. Build what's next. Hugging Face: huggingface.co/stepfun-ai/St… ModelScope: modelscope.ai/models/stepfun… GitHub: github.com/stepfun-ai/Step-5…
3
2
2
21
3,756
Step 5 Preview: The new pareto frontier of agentic finance is here We ran an experiment: which open model gives the best M&A memo per dollar? We had 6 open models write the same memo for 80 companies, and two other AI judges picked the best xitter.cf/StepFun_ai/status/2110…
Step 5 Preview is now open weights. BF16 and FP8 weights are live on Hugging Face and ModelScope, together with our technical blog. • Sparse MoE: 600B total, 27B active per token • 1M-token context window • Native text and vision input • Built for long-horizon agentic work: coding, software engineering, professional knowledge work and finance Frontier intelligence, open weights. Build what's next. Hugging Face: huggingface.co/stepfun-ai/St… ModelScope: modelscope.ai/models/stepfun… GitHub: github.com/stepfun-ai/Step-5…
4
6
1
11
1,204
StepFun retweeted
One of our teammates at StepFun had his X account hacked: The account handle was changed to @ContentCase, and it was used to send fake messages to people on X. The account is currently suspended. I’ve been in touch with him, and @sheriyuo is his temporary new account. We’re still working on recovering the original account and hopefully he’ll be able to get it back and continue to use it Please be careful with any unexpected DMs or messages you may have received from ANY account. Also, take a minute to check your own account security and be extra careful with suspicious links, even when they seem to come from someone or projects you know. Definitely don’t want to see this happen again. If anyone at X can help with the recovery, or knows the right person to contact, please reach out. We’d really appreciate it! @X @StepFun_ai
Chetaslua
@chetasluaOct 1
Phishing attacks through @StepFun_ai verified account please look into this @ContentCase and everyone be safe , they hacked many ai labs employees accounts to do it My review of gemini 4 argon is taking time ( i had bad fever for last two days 😭)
4
4
47
4,762
StepFun retweeted
Tried Step 3.7 Flash in @cline with Playwright MCP. - Select the Step 3.7 Flash model (free) - Installed Playwright MCP through one-click prompt - Let it open the web page, interact with it, and capture screenshots This shows the workflow: efficient agentic model + good tools + fast browser automation. Step 3.7 Flash is free to try in Cline this month. Cheers guys! 😄
2
3
24
4,476
One prompt. One model. A complete fitness app in seconds. 🔥 I tested the new StepFun Step 3.7 Flash and gave it this prompt: "Build a beautiful single HTML fitness tracker app using Tailwind CSS, Chart.js, and Font Awesome. Dark modern theme. 4 tabs Dashboard, Workouts, Diet, Progress. Weekly charts, clickable workout images (Chest/Legs/Back/Full Body), detailed logging, meal tracking with calories & protein, and localStorage." The result? It instantly created a professional-grade Fitness Hub fully working on the first try. ✅ Live stats & weekly progress chart ✅ Clickable workout category images ✅ Add workouts (sets, reps, weight) + delete ✅ Diet logging with meal history ✅ Smooth tabs, animations, and everything saves automatically Opened the single HTML file… and it just worked perfectly. This isn’t just code generation. This is production-ready agent behavior. Step 3.7 Flash is built differently. Who else is trying it? Drop your results 👇 #StepFun #AI #Fitness #OneShotCoding @StepFun_ai
🤝 Paid partnership
27
11
1
80
23,116
StepFun retweeted
Avid
@Av1dliveJun 11
most people stream music in a browser tab buried behind thirty others. eating RAM. serving ads between songs. i wanted to see what happens when you strip music down to pure keyboard control. so i used Hermes + Step 3.7 Flash to build a spotify-style music player directly in the terminal. took about 3 hours. @StepFun_ai instead of a heavy web app, the workflow broke the build into clean modules: → search module that queries the catalog → playback engine that streams audio directly → library manager for playlists and saved tracks → now-playing view that renders album art as ASCII → keybinding layer so you never touch the mouse what stood out was how Step 3.7 Flash handled structure. clear modules. organized logic. the entire terminal-first UX generated in one coherent workflow. i typed a single command and watched it authenticate, pull my playlists, queue an album, and start playing with a live progress bar ; all inside a TUI. zero browser overhead. it feels closer to a native player than anything running in a tab. and this is just the starting point. same pattern extends to: → a self-hosted music server → a CLI wrapper for any streaming API → a developer-focused media player the terminal isn't dead. it's just been waiting for better tooling.
🤝 Paid partnership
13
4
50
5,627
StepFun retweeted
发现一个很离谱的模型,400 TPS 的输出速度,比 DeepSeek-V4 Flash 快 2 倍多 还有超强的多模态理解 + 搜索检索 + 推理综合,这居然是一个 Flash 模型能同时做到的 我让它帮我写了一个竞品分析 Agent: 输入一张产品官网截图,自动完成"读图提取 → 搜索市场数据 → 输出竞争分析"的全流程。 然后用 Cursor 的定价页试了一下,不到一分钟跑完,花了一毛多: 它先读懂了整个页面(四档定价 × 各档功能差异 × 推荐标签与社会证明),提取出完整的结构化数据。 然后自己决定搜什么,自动生成搜索 query 去查竞品定价、用户评价、市场对比。 最后综合视觉提取 + 搜索结果,给了我完整的竞争分析:定价对比、竞品定位差异、真实壁垒在哪里、切入建议。 这个模型就是 @StepFun_ai 的 step-3.7-flash,多模态 + 搜索 + 推理一体化内嵌在模型里,如果你在构建 AI Agent、编码工作流或多模态系统,Step 3.7 Flash 很推荐大家尝试一下: Step海外平台:platform.stepfun.ai​ Step国内平台:platform.stepfun.com
🤝 Paid partnership
52
21
1
133
41,491
StepFun retweeted
我每月 AI 订阅两三千,其中有些是TTS、ASR之类的,主流的几家都挺贵,API协议也都不一样 我一直在想有没有可以一个套餐做到: 语音复刻、会议转写、AI 播客生成、实时语音问答、语音输入、写代码 可算找到了一个活菩萨,阶跃星辰的 Step Plan,一个月6.99美金,根本用不完,就这样,我把其他的陆续停了。 一个订阅里有各类模型,还都是 TOP级: 1、LLM:Step 3.5 Flash,模型延迟低到离谱,也可以接入Claude / Cursor / Cline 2、TTS:stepaudio-2.5-tts(查了一下,排名比 ElevenLabs 高) 3、ASR:实时语音对话,支持音色复刻 4、生图:文生图 + 图像编辑,0.7 秒出图 全部 OpenAI SDK 直连,换一行 base_url。 下面给大家一些案例吧(详情放评论区): 1、英文录音 → 54 秒出中文笔记 2、英文长文 → 双人对谈 mp3 通勤听 3、同一段话 → TTS 演 7 种情绪 4、鲁迅《孔乙己》→ 自动拆角色的有声书 5、英文播客 → 端到端中文重制版 @StepFun_ai
🤝 Paid partnership
96
16
1
144
60,103
StepFun retweeted
很多人卸载了沉浸式翻译,我看上了里面那个贼快的模型供应商 卸载沉浸式翻译之后,用其他翻译插件明显感觉翻译变慢了,是肉眼可见的差异,所以去看了一下它的免费模型供应商,才发现它用的是 Stepfun。 Stepfun 有一个专门面向开发者的订阅计划叫 Step Plan,里面主推的模型是 Step 3.5 Flash 2603,官方定位就是“低延迟推理 + Agent 工作负载”。 这正解释了我之前的体验:不是翻译插件本身写得有多好,而是底层模型的推理管道极短,首 token 延迟压得很低。 为了验证这个结论,我直接买了一个 Step Plan,打算把它接到自己的翻译脚本上,量化看看它到底快在哪。 测试环境很简单,没有用现成的翻译插件,而是自己写了一个极简的网页翻译工具:浏览器扩展抓取当前页面选中文本,通过 API 发给 Step 3.5 Flash 2603,要求直接返回中文译文,不做额外解释。 我选了五个真实的英文网页做测试:一篇 Medium 技术博客(约 800 词)、一段维基百科词条(约 1500 词)、一份 arXiv 论文摘要页(约 3000 词)、再加一篇完整的英文报道(约 10000 词)。 每个网页取整页正文,连续测 10 次取平均值,同时用另一个主流 Flash 级模型做相同任务的对照。 结果可以说非常直观: - 800 词的技术博客,Step 3.5 Flash 2603 首 token 延迟平均 0.17 秒,我几乎感知不到等待。 - 1500 词的维基词条,首 token 延迟 0.20 秒,长段落也稳。 - 3000 词的论文摘要,首 token 延迟 0.25 秒,对比模型在这个量级已经开始出现首 token 延迟跃升到 0.5 秒以上。 - 10000 词的完整报道,这是最考验推理引擎的场景,首 token 延迟 0.34 秒,全程没有限流、没有中断,译文质量也没有因为速度快而出现乱翻或漏段。 对照模型在 3000 词以后的首 token 延迟普遍在 0.5-0.8 秒之间,10000 词时甚至接近 1 秒。 而 Step 3.5 Flash 2603 关闭推理后把“首字延迟”控制在 0.3 秒上下,这个数字正好是大多数人感知“即时反馈”的阈值。 也就是说,从你松手到译文出现,它快得像是提前算好了一样。 @StepFun_ai ⬇️
🤝 Paid partnership
51
18
2
149
44,542
StepFun retweeted
A Chinese AI lab just made running Claude Code dirt cheap It's called Step Plan and it routes between their fastest models so you stop burning $20 every time an agent loops. - Works inside Claude Code, Cursor, Cline, Roo Code, Trae - Step 3.5 Flash built for high-frequency agent calls - StepAudio 2.5 ASR for voice pipelines One subscription. Full toolchain.
🤝 Paid partnership
59
78
200
114,952
Step 3.5 Flash is now open source: model weights and full training framework (SteptronOSS), released together.🚀 196B total, 11B active. SWE-bench Verified 74.4% / Terminal-Bench 2.0 51.0%. - MoE architecture: 288 routed experts + 1 shared, Top-8 activation per token - MTP-3: predicts 4 tokens per forward pass, 100–300 tok/s typical, 350 tok/s peak - 3:1 SWA ratio (1 full attention + 3 sliding window layers): 256K context at lower compute cost - 💻 Runs on Mac Studio M4 Max and NVIDIA DGX Spark - SteptronOSS: SFT, continued pretraining, RL (WIP) - Apache 2.0 Two checkpoints released: Step-3.5-Flash-Base and Step-3.5-Flash-Base-Midtrain. 🤖 Base: modelscope.cn/models/stepfun… 🤖 Midtrain: modelscope.cn/models/stepfun… 🔧 Training Framework: github.com/stepfun-ai/Steptr… 📄 Paper: modelscope.cn/papers/2602.10…
23
54
5
551
40,653
🚀 Meet STEP3-VL-10B—it delivers SOTA-level visual perception, complex reasoning, and human-aligned intelligence—redefining efficiency in open multimodal AI. ✅ Beats or matches models 10–20× larger (like GLM-4.6V, Qwen3-VL, even Gemini 2.5 Pro) ✅ Achieves SOTA on MMMU, MathVision, OCRBench, ScreenSpot, and more ✅ Trained on 1.2T tokens + 1,400+ RL rounds (RLHF + RLVR) ✅ Supports PaCoRe: parallel collaborative reasoning (128K context!) Despite its compact size, it leads the <10B class in: • STEM reasoning (94.43% on AIME 2025 w/ PaCoRe!) • Visual perception (92.05 on MMBench) • GUI understanding & OCR • Spatial reasoning for embodied AI Built with: • PE-lang visual encoder (1.8B) • Qwen3-8B decoder • Multi-crop high-res input (728×728 global + locals) 📥 Base & Chat versions on ModelScope: Base: modelscope.cn/models/stepfun… Chat: modelscope.cn/models/stepfun…
26
114
26
927
217,825
PaCoRe🔥 a new way to scale test-time compute, released by @StepFun_ai ✨ Breaks the context limit with parallel + coordinated reasoning ✨ Reaches million-token TTC without larger context windows ✨ Helps AI think in parallel, not just in long chains ✨ Models, data, inference pipeline all released on @huggingface
6
18
1
106
12,274
GELab-Zero:完全开源的GUI Agent解决方案 集成了模型和基础设施,支持本地部署,提供完整的隐私控制。简化了手机GUI Agent的工程复杂度。能实现查找任务、操作APP等手机端任务。 Github:github.com/stepfun-ai/gelab-…
3
9
1
27
7,916
We are excited to introduce Open Vision Reasoner (OVR) 🚀 — transferring linguistic cognitive behavior to unlock advanced visual reasoning! 💡 Two-stage recipe • Massive linguistic cold-start on Qwen-2.5-VL-7B sparks “mental imagery” • ~1 k-step multimodal RL refines & scales useful patterns 🔍 Insights • Behavior transfer starts surprisingly early • Cold-start memorizes, RL selects & amplifies • Transfer favors high-utility skills like visual reflection 🎯 SOTA scores MathVision 51.8 % MathVerse 54.6 % 🛠️ Code & data dropping soon – stay tuned 🔗 Paper: huggingface.co/papers/2507.0… 🌐 Project: weiyana.github.io/Open-Visio… 🐙 Code: github.com/Open-Reasoner-Zer…
3
27
5
141
13,009
🎉 Introducing Open Reasoner Zero 🚀 Performance: Matches DeepSeek R1-Zero (32B) in just 1/30 steps! 📚 Full training strategies & technical paper 💻 100% open-source: Code + Data + Model ⚖️ MIT licensed - Use it your way! 🌊 Let the Reasoner-Zero tide rise! 🚢 1/n
27
154
15
841
102,176
An upcoming Chinese lab worth keeping an eye on is @StepFun_ai. They released a SOTA audio LLM this week on par with Gemini 3, and just released a new computer use agent which is crushing all benchmarks Find them here: huggingface.co/stepfun-ai
🚀Introducing GELab-Zero: The first complete "Model + Infra" stack for GUI agents. 🤓We are open-sourcing the full stack. Think of it as the open-source answer to GUI Agent MCP. 👯What’s inside: ✅SOTA 4B Model: Lightweight, fast, local execution. ✅One-Click Infra: Plug-and-play. No more ADB or dependency nightmares. ✅AndroidDaily: A new benchmark based on real-world business scenarios. 🤖Performance Highlights: ✅SOTA on Open Benchmarks: Beating larger models (like GUI-Owl-32B) on ScreenSpot, AndroidWorld, and OSWorld. ✅73.4% Accuracy on AndroidDaily, well above UI-TARS-1.5 (47%), Gemini-2.5-pro-thinking (36.6%), and GPT-4o (19.6%). 🔥 Download it, run it locally, break things, improve things. Welcome to GELab-Zero. HuggingFace: huggingface.co/stepfun-ai/GE… GitHub: github.com/stepfun-ai/gelab-… Blog:opengelab.github.io/index.ht…
14
70
3
505
68,247