阶跃星辰正式发布旗舰模型 Step 5 Preview。它采用稀疏 MoE 架构，总参数 600B，每次仅激活 27B，支持 100 万 Token 上下文和图文输入。API…
This is a dev post classified by Jev as AI dev tools (a launch), kept by the Dev Radar because it carries real work, not commentary.
阶跃星辰正式发布旗舰模型 Step 5 Preview。它采用稀疏 MoE 架构，总参数 600B，每次仅激活 27B，支持 100 万 Token 上下文和图文输入。API 和 Studio 已经全量开放，完整模型权重将在 10 月 15 日放出。 此前 Artificial Analysis 已经提前跑完评测。Step 5 Preview 在 Intelligence Index 得到 44 分，与 Kimi K3 Max 持平。AA 当前测得其输出约 100 Token/s，每项任务成本 0.71 美元；Kimi K3 Max 为 2 美元。 官方重点展示了它跑长任务的能力。阶跃让 Step 5 Preview 连续最多 24 小时自主优化一套 H100 GPU 内核。模型会自己改代码、跑测试、比较结果，再继续迭代。约 22 小时后达到 508 TFLOPS，同一实验中的 Claude Opus 5 为 493 TFLOPS。 另一项 24 小时实验里，Step 5 Preview 自己设计后训练数据，把 Qwen3-30B-A3B 在 AIME24 上的准确率从 53.3% 提到 60%，与 Claude Opus 5 持平，同时消耗更少的标注 Token。
Posted by 思维怪怪 (7.8k followers) 17 h ago · 22 likes · 4.1k views · view the original post on X. Kept by the Dev Radar as AI dev tools.
More dev work like this
- Need a second model for a Codex subtask — without changing your controller? — @DanKornas
- This will never get old. — @blizaine
- We reverse-engineered Instinct's memory system and implemented it with supermemory's… — @supermemory
- Grok @bot is now my personal app developer. It’s bringing my 6 year old iOS code up to… — @IowaTesla
- You can run locally Google's Nano Banano 2 level Image model Qwen-Image-2.1 on RTX 3090… — @0x0SojalSec
- Qwen just released Qwen-Image-2.1, and this is exactly why open models matter. — @aijoey
- Hahaha sooo cool.. have intermittent but continuous context compaction working.. that… — @u1tra_instinct
- When the cost of error is high, trust is everything. — @github
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 16.1k posts from 4.9k X accounts over the last 21 days, 1.8k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-20 19:24 UTC. Full method.