OpenCode Unlocks Step 5 Preview for 1 Week Free: StepFun 600B Flagship MoE, 1M Context, and Zero Data Retention

OpenCode Unlocks Step 5 Preview for 1 Week Free: StepFun 600B Flagship MoE, 1M Context, and Zero Data Retention

Following several waves of anonymous testing models and temporary compute allocations, OpenCode has officially announced a new free benefit: the next-generation flagship model Step 5 Preview is now available on OpenCode for a one-week free public trial for all users.

Unlike previous anonymous tests that sparked community speculation such as Ox Alpha and Space Bunny Alpha, Step 5 Preview arrives without a mystery mask. It is the new flagship foundation model unveiled in late September by StepFun. Built on a sparse Mixture-of-Experts architecture featuring 600B total parameters and 27B active parameters, the model natively supports a 1M token context window and multimodal inputs, fine-tuned specifically for software engineering and long-horizon agentic workflows. Crucially, both the platform and the provider explicitly commit to Zero Data Retention, providing solid guarantees for developer privacy.

With the official open-source weights release date approaching, this one-week zero-cost trial offers developers an ideal window to benchmark its real-world engineering capabilities. This guide provides a quick breakdown of Step 5 Preview's core specifications, integration paths, and practical tips.

💡 Key Highlights

• ⚡ Step 5 Preview Free for 1 Week: Starting October 8, 2026, all users enjoy 7 days of 0-cost access with zero credit or balance deduction.
• 🧠 StepFun 600B Flagship MoE Architecture: 600 billion total parameters with 27 billion dynamically activated per token, balancing top-tier deep reasoning with fast inference throughput.
• 📜 1M Context Window and Native Multimodal Input: 1 million token context capacity supporting text, image, and video inputs, effortlessly digesting entire repositories or voluminous specifications.
• 🛡️ Zero Data Retention: Both the platform and model provider commit that sensitive code and prompts are never stored, logged to disk, or reused for downstream training.
• 🛠️ Seamless Multi-Platform Setup: Directly switchable in OpenCode CLI and desktop apps, or instantly accessible via standard OpenCode Zen APIs across Cursor, Claude Code, Codex, Pi, and ZCode.
• ⏳ Sneak Peek Before Open-Source Release: StepFun is scheduled to release full open-source weights on October 15, 2026, making this limited trial a valuable preliminary evaluation opportunity.

🔥 1. Event Details and Core Model Specifications

While some developers initially assumed this was another anonymous blind test from OpenCode, the announcement clearly identifies the model as Step 5 Preview. Developers familiar with leading Chinese AI labs will recognize its pedigree:

1. 600B Sparse MoE and the Pareto Frontier

Step 5 is StepFun's flagship foundation model purpose-built for real-world agentic execution and complex knowledge work. Its primary architectural highlight is the sparse MoE structure pairing 600B total parameters with 27B active parameters.

This setup represents a Pareto-frontier tradeoff: it delivers the deep abstraction and broad domain intelligence typical of massive 600B models while activating only 27B parameters per token, maintaining responsive latency and cost-effective computational throughput. In independent open-source evaluations, Step 5 Preview consistently ranks among the global top tier.

2. 1M Context Window and Multimodal Capabilities

In day-to-day software engineering, a 1M token window offers tremendous operational flexibility:

• Full-Repository Understanding: Ingest whole project directory structures, complex dependency graphs, and extensive API specifications in a single prompt without fragmentation;
• Native Visual Inspection: Upload error screenshots, architecture wireframes, or UI design mockups directly, allowing the model to diagnose CSS styling collisions or layout anomalies with full visual context;
• Long-Horizon Agent Stability: Tuned specifically for multi-step reflection and chained Tool Calling, noticeably reducing attention drift across long task sequences.

3. Enterprise-Grade Zero Data Retention

When routing proprietary codebases through third-party platforms, privacy and compliance are paramount. OpenCode and StepFun have made a clear Zero Data Retention pledge: inference requests are processed in memory and discarded immediately, never persisted to disk or recycled into future fine-tuning pipelines.

🛠️ 2. Quickstart and Configuration Guide

During this week-long free window, you can choose whichever workflow best fits your habits:

1. Out-of-the-Box via OpenCode CLI and Desktop

If you already use OpenCode:

• Terminal CLI: Run opencode in your shell, type /models to open the selection list, and choose step-5-preview;
• Desktop Application: Press Ctrl + T to summon the model switcher, search for step-5-preview, and start chatting immediately.

2. Connect Mainstream IDEs via OpenAI-Compatible API

If you prefer coding inside your dedicated editor, connect through the OpenCode Zen API gateway. Zen adheres to standard OpenAI API specifications and works seamlessly with Cursor, Claude Code, Codex, Pi, and ZCode:

• Base URL: https://opencode.ai/zen/v1
• API Key: Your personal OpenCode API key
• Model Name: step-5-preview

Requests dispatched to this endpoint during the promotional week are completely free of charge.

⚠️ 3. Practical Tips and Best Practices

To get the most value out of this model during the trial week, keep the following suggestions in mind:

1. Manage the 7-Day Window: The promotion launched on October 8, 2026, and is expected to conclude around October 15. Prioritize backlog tasks like monolithic code refactoring, legacy dependency audits, or architecture documentation during this period.
2. Use 1M Context Judiciously: While 1M tokens provide vast headroom, maxing out context length naturally increases time-to-first-token (TTFT). For single-function fixes or straightforward bugs, keep prompts focused on relevant files and reserve giant context payloads for global refactoring.
3. Scale Visual Inputs Reasonably: When submitting mockups or error captures, crop or downscale images to 1080p. This preserves legible typography while avoiding unnecessary visual encoding latency.

💡 4. Summary and What to Watch Next

OpenCode's free rollout of Step 5 Preview provides substantial zero-cost compute while offering a valuable preview of upcoming open-source technology.

StepFun has scheduled the release of open weights for October 15, 2026. Taking advantage of this free week to test Step 5 Preview against your private repositories and daily agentic loops will help your team determine whether self-hosting or custom fine-tuning makes sense once weights drop.

Take advantage of this limited-time free access and throw your most challenging code tasks at Step 5 Preview today.

📬 Follow Free AI API

To stay ahead of limited-time trials, free API allocations, and model price drops across top AI providers, follow freeaiapi.org. We continuously track high-value compute resources to bring you verified, reliable, and credit-card-free developer perks.