Step 5 Preview 600B: 27B active MoE, 1M context

StepFun released Step 5 Preview as open-weights, providing developers with a self-hostable sparse Mixture-of-Experts model for long-context agentic work. The model contains 600 billion total parameters with 27 billion active per token, supports a context window of up to 1 million tokens, and is now available on Hugging Face and ModelScope in FP8 and BF16 precision formats. The September 20 API launch preceded the October 14 weights release.

What changed

The open-weights release on October 14 follows StepFun’s commercial API launch earlier in September. Developers can now download and deploy Step 5 Preview on their own data-center infrastructure using either FP8 or BF16 precision. The FP8 checkpoint weighs 650.32 GB. Dedicated Docker images compatible with vLLM and SGLang provide OpenAI-compatible API interfaces for deployment.

Step 5 Preview uses a narrow, deep architecture comprising 92 transformer blocks. The sparse design activates only 27 billion parameters per token from the 600 billion total, maintaining computational efficiency during inference. The model combines gated sliding-window attention with full attention layers in a 3-to-1 ratio. Its sparse grouped-query attention mechanism selects 4,096 history tokens per query, scoring them in blocks of eight.

The model includes speculative decoding via an MTP-3 head that reportedly provides approximately 30 percent end-to-end speedup. StepFun trained Step 5 Preview using FP8 mixed precision and reinforcement learning on million-token verifiable agentic environments.

Why it matters

Step 5 Preview’s open-weights release provides teams building long-horizon agentic systems an alternative to proprietary frontier models for self-hosted deployment. Organizations working on software engineering, professional knowledge work, or financial analysis can now run the model on their own hardware without relying on API calls. The million-token context window enables agentic systems to reason over large codebases, process multiple documents simultaneously, and coordinate complex multi-file interactions.

This release represents a trade-off between API convenience and infrastructure control. Unlike managed services, running Step 5 Preview requires data-center GPU resources and operational overhead but eliminates dependency on external services. The model occupies a position between smaller open-source alternatives and larger proprietary systems, sharing architectural philosophy with other sparse MoE releases like Meta’s Muse Spark Open.

StepFun maintains a commercial API as an alternative for teams preferring managed infrastructure. The API pricing structure charges $1.00 per million input tokens and $2.70 per million output tokens. Cache hits cost $0.10 per million tokens, with output pricing including reasoning tokens.

What to test

Before deploying Step 5 Preview for production agentic work, evaluate these factors against your specific requirements.

Benchmark performance context

StepFun reported results using the High reasoning setting while comparing against competing models evaluated at Max effort, creating incompatible comparisons. Verify actual performance on your target tasks rather than relying solely on published benchmarks. Independent evaluations from Artificial Analysis show Step 5 Preview scoring 88.3 percent on AA-LCR long-context reasoning and 46.5 percent on Humanity’s Last Exam. However, significant weaknesses appear: 0.4 percent on SRE Bench and 9.2 percent on Harvey’s Legal Agent Benchmark. The model achieved 67.7 percent on DeepSWE and 33.3 percent on Terminal-Bench 4.0. Conduct your own testing on representative datasets before production deployment.

Infrastructure requirements and scaling

Confirm your available hardware matches the stated requirements. FP8 deployment requires either 4 NVIDIA B300 GPUs or 8 H200 GPUs on a single node. BF16 deployment requires 8 B300 GPUs on one node or 16 H200 GPUs distributed across two nodes. Measure inference latency and throughput with your expected query patterns and batch sizes. StepFun recommends temperature 1.0, top_p 0.95, and top_k 0, though your application may require different hyperparameter values.

Reasoning token economics

The API charges for reasoning tokens as output costs. Evaluate whether the reasoning overhead justifies quality improvements for your workload. Test how low, medium, and high reasoning effort settings balance quality and speed for your use case.

Cost-benefit analysis

Compare API pricing against the operational costs of self-hosting on your infrastructure. High-volume production systems may achieve lower per-token costs through self-hosting after accounting for infrastructure investment, but self-hosting introduces operational complexity and maintenance requirements.

The conclusion

Step 5 Preview open weights provides a practical option for teams requiring self-hosted agentic systems with strong long-context capabilities and sparse efficiency. The model demonstrates strong performance on long-context reasoning with 88.3 percent on AA-LCR and achieves 67.7 percent on DeepSWE, making it suitable for software engineering agents and financial analysis workflows. Performance gaps in legal reasoning and SRE tasks require careful evaluation if those domains are critical to your system.

The self-hosted model does not match performance of frontier proprietary systems like GPT-6 Astra or Claude Opus 5 across most benchmarks. It competes instead with other open-weight alternatives such as GLM-5.3 and Kimi K3, where Step 5 Preview shows competitive performance on coding and finance tasks despite lagging on specialized domains.

For teams managing data-center infrastructure and needing reliable long-context agentic capabilities, the open-weights availability provides a pragmatic deployment path. Teams prioritizing simplicity can utilize the commercial API instead. Both approaches require validation on your specific workload before full production deployment. Monitor StepFun’s future model releases and emerging independent benchmarks on agentic tasks, as the demonstrated long-context reasoning strengths position this model favorably as agent systems advance.

Related reading

• Beam 501B: 23B active parameters, Apache 2.0 licensed
• Apertus 1.5: 262K context, image and audio support
• QwenGyre elastic RL framework trains long-horizon agents