Step 5 Preview: 600B MoE Model Brings 27B Active Parameters and 1M Context

By AiCybr Editorial · Published September 21, 2026

StepFun has introduced Step 5 Preview, a flagship sparse Mixture-of-Experts model built for agentic software engineering and professional knowledge work. The model has 600 billion total parameters, 27 billion active parameters per token, a 1 million-token context window and vision input, according to StepFun's September 2026 technical release.

The architecture makes Step 5 Preview substantially larger in stored parameter count than StepFun's earlier Flash models while keeping per-token activation far below the full 600B checkpoint. StepFun is positioning the model for long-running agent workflows that combine reasoning, tool use, code execution, research and artifact creation.

StepFun reports a 67.7 score for Step 5 Preview at High reasoning on DeepSWE v1.1. In the company's comparison, Kimi K3 Max scores 67.5, GLM-5.3 Max scores 66.9, GPT-6 Astra Max scores 74.1 and Claude Opus 5 Max scores 74.0. These are vendor-presented benchmark results and should be interpreted within the published evaluation setup.

Step 5 Preview specifications

The 27B active-parameter figure describes inference routing, not checkpoint size. Deployment memory for an eventual downloadable checkpoint would depend on the full model weights, precision, quantization, KV cache and serving implementation. StepFun's current launch material focuses on the hosted preview and model behavior, so local hardware sizing should wait for an actual weight release and serving specification.

Coding and agent benchmarks

StepFun reports an avg@4 score of 49.0% on StepCodeBench, with the model performing particularly well in bug repair, feature modification and refactoring. The company says performance approaches Claude Opus 5 on low- and medium-difficulty tasks, while its own evaluation shows a larger gap on the most difficult long-running tasks.

The DeepSWE result provides a second coding signal. Step 5 Preview High reaches 67.7 in StepFun's published comparison, close to Kimi K3 Max at 67.5 and GLM-5.3 Max at 66.9, while GPT-6 Astra Max and Claude Opus 5 Max remain above 74 in that table.

Benchmark scores remain sensitive to harnesses, reasoning settings, tool access and evaluation revisions. The useful comparison is therefore the complete evaluation configuration and workload, especially for agent models that can consume tools and large token budgets during a task.

A 1M-token context targets long-running work

The 1 million-token context window is one of the model's most consequential deployment characteristics. Agentic coding and research systems can accumulate repository files, tool outputs, execution traces, documents and earlier decisions over long sessions. A larger context budget can reduce the frequency of aggressive summarization or state eviction when the application actually supplies that material.

StepFun's demonstrations emphasize sustained execution. In one post-training experiment, the model selected how to use an annotator, revised training data and tracked downstream performance. The resulting model reached 60% on the official AIME24 test, up from 53.3% before post-training, according to StepFun.

A separate Pokémon Red run exceeded 3,000 turns and 6 million cumulative interaction tokens. By turn 3,082, StepFun reports that the agent had obtained Cut, earned three Gym Badges and defeated Lt. Surge. The cumulative token figure spans the ongoing interaction rather than representing a single 6M-token prompt; the model's stated context window remains 1M tokens.

Tool use extends beyond coding

StepFun also demonstrates the model operating against programmable hardware when given documentation and user authorization. The described interfaces include cameras, COM ports, screenshots and simulated mouse input, allowing a coding agent to connect software changes with observable hardware behavior.

For research workloads, StepFun reports a climate-analysis task covering 1,000 locations over 25 years. The agent coordinated 950 web fetches in one action and assembled 300,000 monthly records across 11 variables before analyzing regional solar-seasonality patterns.

Another demonstration produced a 17-sheet analytical workbook with 11,057 rows, formulas, charts, source data and regional panels for a diesel-surcharge review. These examples are demonstrations of workflow breadth rather than standardized cross-model benchmarks, but they show the type of long-horizon artifact work StepFun is targeting.

What the 600B/27B design means for inference

Sparse MoE models route each token through a subset of their available experts. Step 5 Preview's 600B total and 27B active figures therefore separate model capacity from per-token activated compute.

That distinction can improve inference economics relative to a hypothetical 600B dense model, although active parameter count alone does not determine real serving cost. Memory bandwidth, expert routing, interconnect traffic, KV-cache size, context length, batching, quantization and kernel quality all affect throughput and latency.

The 1M context ceiling can also create substantial KV-cache and attention costs when applications approach the limit. Production evaluation should measure the context lengths that real workloads use instead of sizing infrastructure from the maximum advertised window alone.

Availability and validation priorities

Step 5 Preview is presented by StepFun as a preview model. The strongest technical claims currently come from StepFun's own launch material, including its architecture figures, DeepSWE and StepCodeBench results, and long-horizon demonstrations.

Independent evaluation should focus on coding-agent reliability across multiple harnesses, token consumption per completed task, long-context quality at increasing prompt lengths, tool-call accuracy, latency and total API cost. Those measurements determine whether the model's large context and sparse architecture translate into better production economics for a particular agent workload.

Bottom line

Step 5 Preview expands StepFun's agent-model line to a 600B-total, 27B-active sparse MoE architecture with vision and a 1M-token context window. Its published DeepSWE score of 67.7 and StepCodeBench avg@4 result of 49.0 place it in a competitive coding-agent tier within StepFun's evaluation set.

The model is most relevant to teams evaluating long-running coding, research and professional-work agents. The next decision-grade measurements are independent task completion, latency, token consumption and serving economics across realistic agent harnesses.

Sources

• StepFun — Step 5 Preview technical launch: https://www.stepfun.com/step-5-preview
• Independent launch/specification cross-check, CellCog: https://cellcog.ai/blog/step-5-preview/