k
keynews.ai
keynews.ai
FrontierThe story, in brief

StepFun Launches Step 5 Preview: A 600B-Total, 27B-Active MoE Model With 1M Context for Long-Horizon Agentic Work

600B parameters, 27B active, 1M context. StepFun's Step 5 Preview targets long-horizon agentic work at $1/1M input tokens.

Reporting by MarkTechPost 1 min brief
Illustration of a transparent lens revealing connected networks across layers of paper.
Exploring the next frontier of AI research.AI illustration by KeyNews
The KeyNews take

Why it matters

StepFun enters the sparse MoE race with a large-context model positioned for agent and knowledge-work use cases. The pricing and multimodal capabilities signal competitive pressure in the frontier lab space, and the October open-weight timeline affects the open-model landscape.

The key facts

7 to know

  1. 600B total parameters, 27B active per token (sparse MoE architecture)

  2. 1M token context window

  3. Multimodal: text, image, video input

  4. API pricing: $1.00 per 1M input tokens, $2.70 per 1M output tokens

  5. Open weights scheduled October 15, 2026

  6. Target use cases: software engineering, professional knowledge work, finance, long-horizon agentic work

  7. API access live now (as of Sep 2026)

Go to the source

MarkTechPostmarktechpost.com

Publisher excerpt: StepFun has released Step 5 Preview, a sparse Mixture-of-Experts model with 600B total parameters and 27B active per token. It supports a 1M-token context window and accepts text, image, and video input. The model targets long-horizon agentic work in software engineering, professional knowledge…
Read original report
Back to today's editionMore frontier news

The wider picture

View all
Illustration of a transparent lens revealing connected networks across layers of paper.
AI illustration by KeyNews
Frontier01

Alibaba's open-weight Qwen-Image-2.1 claims to beat closed models in image generation with just 7 billion parameters

A capable open-weight image model at 7B parameters challenges the closed-model dominance in generation and editing, expanding practitioner options for on-device and cost-efficient image workflows.

The Decoder
Illustration of a transparent lens revealing connected networks across layers of paper.
AI illustration by KeyNews
Frontier02

Tencent's Gander aims to keep talking while it works in the background

A novel architecture for multimodal agents that separates conversational continuity from task execution. Demonstrates a real capability tradeoff: smoother UX vs. task reliability. Relevant to how frontier labs are rethinking agent design.

The Decoder
Illustration of a transparent lens revealing connected networks across layers of paper.
AI illustration by KeyNews
Frontier03

Simulated students that make realistic mistakes help AI tutors learn faster

A novel approach to AI training using realistic synthetic feedback loops is accelerating tutor model development and reducing the cost of evaluation data. This represents a meaningful shift in how frontier labs can iterate on capability without massive labeled datasets.

The Decoder