Step

Overview

Step AI is a multimodal AI assistant launched by Step Star, based on a trillion-parameter model. The company was founded in April 2023 by Jiang Daxin, former Microsoft Global Vice President, with a valuation exceeding $17 billion, making it one of the leading AI financing companies in China. In May 2026, it released the StepAudio 2.5 real-time voice model, scoring 80.41 in paralinguistic perception (tone, pauses, chuckles, etc.), surpassing GPT-Realtime.

Major Strategic Upgrade on 2026-07-13: Step Star held a press conference in Shanghai, unveiling a trio of products—the world's first Agent-native Operating System Step AOS (Step Agentic-native OS, deeply restructuring hardware scheduling and system capability allocation to break down the three walls of memory, decision-making, and action), the large model-native AI terminal brand STEPX, the next-generation personal agent Step Amoo built on the model matrix and Step AOS, and the world's first large model-native agentic phone STEPX Neo. This marks Step Star's completion of the full technical loop from "foundation model → agent system → hardware terminal," extending from a pure software vendor to a full-stack "model + system + terminal" player, echoing ByteDance's recent move into agentic phones. The industry widely interprets this as a key step for large model companies to find a sustainable path for "terminal-driven commercialization."

Foundation Model and Speech Updates: On 2026-09-20, StepFun launched the flagship Step 5 Preview, built on a sparse MoE architecture with 600B total and 27B activated parameters, a 1M-token context window and text plus vision input. It scored 44 on the Artificial Analysis Intelligence Index, and StepFun states a per-task cost of one eighth that of Claude Opus 5. Weights are scheduled for open-source release on 2026-10-15. On 2026-09-22, Artificial Analysis evaluated Step 5 Preview with an Intelligence Index of 44 and a per-task cost of roughly 1/2.8 of comparable models, making cost efficiency the main point of this tier. On 2026-10-09, StepFun announced Step 5 Preview is free for one week on Nous Portal and Kilo; the model is a 600B-parameter MoE with 27B activated per token, supports a 1M token context and vision input, scores 33.89 on the Hermes Index, and targets coding and finance performance at lower cost. On 2026-09-15, StepFun released five speech models in the StepAudio 3 series — Realtime, ASR, TTS, Gen and Music — all now available on its open platform. Realtime supports native full-duplex real-time conversation and ranks first globally on the Artificial Analysis Conversational Dynamics leaderboard with a composite score of 98.9%, while also placing first on the Speech Reasoning leaderboard with 99.7% word-level speech reasoning accuracy; for complex questions it supports parallel reasoning and speech generation, plus Tool Call and asynchronous execution of long tasks. ASR ties for first globally with a 1.7% non-streaming word error rate. TTS covers paralinguistic cues such as laughter, hesitation, stuttering and self-correction, and supports streaming generation. Gen unifies the generation of vocals, sound effects, ambient audio and background music, while Music supports multi-turn composition based on ABC notation.

Key Features

• Step AOS Agent-native Operating System (New): Released on 2026-07-13, the world's first operating system rebuilt for agents, deeply restructuring hardware scheduling and system capability allocation to break down the three walls of memory, decision-making, and action
• STEPX Neo Agentic Phone (New): The world's first large model-native agentic phone, powered by Step AOS, integrating model, system, and terminal
• Step Amoo Personal Agent (New): A next-generation personal agent built on the Step model matrix and Step AOS, serving as the core interaction layer for STEPX terminals
• Step 5 Preview Foundation Model (New): Launched 2026-09-20, a sparse MoE flagship with 600B total and 27B activated parameters, a 1M-token context window and text plus vision input; scored 44 on the Artificial Analysis Intelligence Index, with a per-task cost of one eighth that of Claude Opus 5 and weights scheduled for open-source release on 2026-10-15
• StepAudio 3 Series (New): Released 2026-09-15 with five speech models — Realtime, ASR, TTS, Gen and Music — all available on the open platform; Realtime ranks first globally on the Artificial Analysis Conversational Dynamics leaderboard at 98.9% and first on the Speech Reasoning leaderboard at 99.7% word-level accuracy, ASR ties for first globally with a 1.7% non-streaming word error rate, TTS covers paralinguistic cues such as laughter, hesitation, stuttering and self-correction with streaming generation, Gen unifies vocals, sound effects, ambient audio and background music, and Music supports multi-turn composition based on ABC notation
• Multimodal Interaction: Supports text, image, voice, and other input methods
• Real-time Voice Conversation: StepAudio 3 Realtime supports native full-duplex real-time conversation, with parallel reasoning and speech generation for complex questions, plus Tool Call and asynchronous execution of long tasks
• Code Generation: Strong programming assistance, supporting multi-language code generation
• Long Text Processing: Supports ultra-long context understanding and processing, with Step 5 Preview offering a 1M-token context window

Use Cases

• Everyday AI conversations and Q&A
• Real-time voice interaction scenarios
• Code development assistance
• Creative content writing
• Multimodal content understanding
• Early adopters and product professionals interested in the new "agentic phone" terminal form
• Coding and finance tasks with Step 5 Preview
• Speech generation, transcription and music composition with the StepAudio 3 series

Pros

• Trillion-parameter model with comprehensive capabilities
• Excellent StepAudio voice model experience, globally leading paralinguistic perception
• First to complete the full-stack model + system + terminal loop; Step AOS is the world's first agent-native operating system
• $17 billion valuation, technical strength recognized by capital
• Balanced capabilities in conversation, creation, and coding
• Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index with a per-task cost of one eighth that of Claude Opus 5, and weights are scheduled for open-source release on 2026-10-15
• StepAudio 3 Realtime ranks first globally on the Artificial Analysis Conversational Dynamics leaderboard at 98.9% and first on the Speech Reasoning leaderboard at 99.7% word-level accuracy, while ASR ties for first globally with a 1.7% non-streaming word error rate

Pricing

Step AI App/web version offers basic features for free, with advanced features paid. Step 5 Preview is free for one week on Nous Portal and Kilo (announced 2026-10-09). Pricing for the STEPX Neo phone hardware and Step AOS system is subject to official press conference announcements (specific retail prices were not disclosed at the press conference; please follow the official website for future launch information).

Summary

Step AI is a top-tier domestic large model product, with voice interaction as its differentiated highlight. The strategic upgrade on 2026-07-13, is even more noteworthy—Step Star is one of the first domestic companies to extend large model capabilities to operating systems and hardware terminals. The combination of Step AOS + STEPX Neo represents an early exploration of the new "agent-native" terminal paradigm, echoing ByteDance's simultaneous entry. The foundation model line has since advanced with Step 5 Preview (2026-09-20), a sparse MoE flagship with 600B total and 27B activated parameters, a 1M-token context window and text plus vision input, scoring 44 on the Artificial Analysis Intelligence Index and stating a per-task cost of one eighth that of Claude Opus 5, with weights scheduled for open-source release on 2026-10-15; Artificial Analysis (2026-09-22) measured a per-task cost of roughly 1/2.8 of comparable models, and from 2026-10-09 the model is free for one week on Nous Portal and Kilo. The StepAudio 3 series (2026-09-15) brings Realtime, ASR, TTS, Gen and Music to the open platform, with Realtime ranking first globally on the Artificial Analysis Conversational Dynamics leaderboard at 98.9% and first on the Speech Reasoning leaderboard at 99.7% word-level accuracy, and ASR tying for first globally with a 1.7% non-streaming word error rate. Suitable for users needing multimodal AI assistants, and for product professionals tracking the "next-generation AI terminal form."

Version History

• Step 5 Preview opens free on Nous Portal and Kilo (2026-10-09): StepFun announced Step 5 Preview is free for one week on Nous Portal and Kilo. The model is a 600B-parameter MoE with 27B activated per token, supports a 1M token context and vision input, scores 33.89 on the Hermes Index, and targets coding and finance performance at lower cost.
• Step 5 Preview scores 44 on Artificial Analysis at about 1/2.8 the cost of peers (2026-09-22): Artificial Analysis evaluated StepFun’s Step 5 Preview: an Intelligence Index of 44 with a per-task cost of roughly 1/2.8 of comparable models, making cost efficiency the main point of this tier.
• StepFun launches flagship Step 5 Preview, weights to be open-sourced on October 15 (2026-09-20): StepFun released Step 5 Preview, its flagship foundation model, built on a sparse MoE architecture with 600B total and 27B activated parameters, a 1M-token context window and text plus vision input. It scored 44 on the Artificial Analysis Intelligence Index, and StepFun states a per-task cost of one eighth that of Claude Opus 5. Weights are scheduled for open-source release on October 15.
• StepAudio 3 series launches, with Realtime and ASR ranking first globally on the Artificial Analysis leaderboard (2026-09-15): StepFun releases five speech models in the StepAudio 3 series — Realtime, ASR, TTS, Gen and Music — all now available on its open platform. Realtime supports native full-duplex real-time conversation and ranks first globally on the Artificial Analysis Conversational Dynamics leaderboard with a composite score of 98.9%, while also placing first on the Speech Reasoning leaderboard with 99.7% word-level speech reasoning accuracy; for complex questions it supports parallel reasoning and speech generation, plus Tool Call and asynchronous execution of long tasks. ASR ties for first globally with a 1.7% non-streaming word error rate. TTS covers paralinguistic cues such as laughter, hesitation, stuttering and self-correction, and supports streaming generation. Gen unifies the generation of vocals, sound effects, ambient audio and background music, while Music supports multi-turn composition based on ABC notation.
• Step AOS + STEPX Brand + STEPX Neo (2026-07-13): Shanghai press conference launched the world's first agent-native operating system Step AOS, terminal brand STEPX, personal agent Step Amoo, and the world's first large model-native agentic phone STEPX Neo, completing the model-system-terminal full-stack loop
• StepAudio 2.5 (2026-05): Released real-time voice model, scoring 80.41 in paralinguistic perception (tone/pauses/chuckles), surpassing GPT-Realtime