The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
The Release Is About Longer Task Chains, Not Simply a Larger Model
StepFun has released Step 5 Preview, a sparse Mixture-of-Experts model aimed at agentic work in software engineering, professional knowledge work, and finance. It has roughly 600 billion total parameters and activates about 27 billion per token. The model accepts text, images, and video, produces text output, and exposes low, medium, and high reasoning effort together with tool calling, JSON Mode, JSON Schema, streaming, and prompt caching.
Those specifications target workflows that repeatedly read material, call tools, run code, and process tool returns rather than ordinary question answering. StepFun says the model coordinated 950 web fetches in a single agent action on a research task. That claim shows the intended operating loop, but it does not establish that all 950 fetches were completed reliably or correctly. For a technical leader, the central question is therefore shifting from how well the model answers one prompt to whether it can preserve direction, cost control, and error handling across a long stateful workflow.
Sparse Computation Saves Per-Token Compute, Not the Whole Machine
Step 5 Preview uses an MoE design in which roughly 4.5 percent of the parameters are active for each token. This can reduce computation per token while preserving a much larger total capacity. StepFun also uses a narrow-and-deep 92-layer Transformer rather than continuing to widen the network. The research team’s argument is that more layers create a longer path for implicit multi-hop reasoning, especially when an agent is processing an expanding stream of tool results and long prefixes.
However, “27B active” cannot be translated directly into the deployment requirements of a 27B model. The full 600 billion parameter pool still has to be available in the serving memory system. The material’s simple estimate puts the BF16 weights at about 1.2 TB before accounting for the KV cache, which means self-hosting will likely require multi-GPU server hardware once the weights are released. A one-million-token context also moves pressure into caching, prefill, scheduling, and long-request concurrency. It does not disappear because the architecture diagram is sparse.
The Price Looks Aggressive Until Extended Reasoning Changes the Equation
StepFun currently offers access through a hosted API and its platform. The listed prices are $1.00 per million uncached input tokens, $0.05 per million cached input tokens, and $2.70 per million output tokens. The input and cache rates create meaningful room for long-context agents, particularly when the same codebase, research corpus, or task state is reused across turns. For teams that have not yet established workflow value, an API is also a lower-friction experiment than buying hardware immediately.
Yet an agent’s cost cannot be modeled from input tokens alone. The material says that output pricing includes reasoning tokens, and Artificial Analysis recorded 160 million output tokens from Step 5 Preview in its Intelligence Index run, compared with a 92 million median. That test does not represent every production task, but it exposes an important risk: low unit prices can be offset by high reasoning effort, retries, tool calls, and runaway loops. Prompt caching reduces repeated-prefix cost, but it does not decide when a task should stop or prevent an incorrect result from contaminating the agent’s state.
Long-Horizon Reinforcement Learning Is a System Capability and an Evaluation Burden
StepFun emphasizes on-policy, long-horizon reinforcement learning and describes bit-wise alignment between MoE routing during training and inference. The material also lists MTP-3 speculative decoding, FP8 MoE, and KV-cache offload, while claiming more than a threefold end-to-end speedup for long-horizon reinforcement learning. Together, these choices suggest that Step 5 Preview is intended to reduce breaks in the loop of searching, executing, reading returns, and planning again rather than merely maximizing one benchmark score.
Long tasks are also harder to evaluate and attribute. The launch material reports 67.7 on DeepSWE v1.1, 49.0 on StepCodeBench, and 80.5 on ProgramBench. StepFun also describes two 24-hour agent experiments, including tuning an H100 kernel to 508 TFLOPS and raising Qwen3-30B-A3B on AIME24 from 53.3 percent to 60 percent through automated post-training. These are company-reported results, and some competitors were run at different reasoning settings, so they should not be treated as an independent like-for-like ranking. Artificial Analysis gives the model an Intelligence Index score of 44, a useful reminder that long-horizon execution, benchmark performance, and general intelligence are different dimensions.
For Production Teams, Validate the Workflow Before Migrating the Stack
The most practical path for Step 5 Preview is to treat it first as an API experiment rather than immediately making “one million tokens of context” an architectural commitment. The first workflows to test should have explicit success conditions, such as modifying a codebase, producing an evidence-backed conclusion after research, or coordinating several tool stages. Teams should log completion rate, tool-call count, output tokens, retry rate, cache hits, and human takeovers at the same time. Otherwise, they may see the unit price while missing the real cost per completed task.
Self-hosting is a separate decision. Open weights are scheduled for October 15, 2026, so hardware requirements, quantization behavior, and inference throughput should not be treated as validated before release. Even if the weights arrive on schedule, roughly 1.2 TB of BF16 weights, a million-token context cache, and a 600-billion-parameter serving stack create separate questions around cost, failure domains, concurrency, and data governance. A workable judgment is to establish an end-to-end baseline through a controlled API when the value comes from longer action chains. Consider self-hosting only when traffic, data residency, or latency requirements show that hosted access is inadequate, and treat it as an infrastructure project rather than a model toggle.