StepFun opens 600B-parameter Step 5

A very large model just became downloadable

StepFun has opened the weights for Step 5 Preview, its flagship model for agentic work, turning what had been a hosted-model story into a self-hosting story for teams with serious infrastructure budgets [1]. The headline numbers are deliberately attention-grabbing: 600 billion total parameters, 27 billion active parameters per token, and a context window of up to one million tokens [1]. In other words, Step 5 Preview is not a small open model aimed at laptops; it is a sparse Mixture-of-Experts system built to make very large capacity available while activating only a fraction of the network for each token [1].

The weights are live through Hugging Face and ModelScope, with reporting identifying FP8, BF16 and NVFP4 checkpoint variants in the official StepFun namespace [3]. MarkTechPost’s October 16 write-up also describes FP8 and BF16 checkpoints as live on the two model hubs, positioning the drop as a new self-hostable option for coding agents, document analysis and finance workflows [2]. AI Tamer’s check of the repositories adds a useful timestamp detail: Hugging Face’s API showed last-modified times for the three official repositories on October 15, 2026, between 01:26 and 01:31 UTC [4].

This is the material change in the story. Step 5 Preview had already been discussed as a hosted model; now developers can inspect, download, deploy and adapt official weights rather than simply call a closed endpoint [3]. That matters because the next phase of agent competition is increasingly about custom workflows: codebases, tool chains, private documents, internal finance models, and long-running tasks that companies may not want to route entirely through third-party APIs.

What Step 5 Preview is built for

StepFun’s own model card calls Step 5 Preview a flagship model for agentic work, with strengths in software engineering, professional knowledge work and finance [1]. The model is described as supporting repository-level software engineering, repeated tool use, long-document analysis, charts, screenshots, images and structured reports [1]. The core bet is that long-horizon agents do not just need smarter next-token prediction; they need enough working context to carry plans, files, traces, logs and references through a multi-step process.

That is where the one-million-token window becomes the marquee feature. A context of that size can hold multiple books, large code repositories, substantial legal or financial document sets, or an entire chain of intermediate agent observations. Or, less solemnly, it could swallow several Dune appendices before breakfast. The serious point is that long-context models change the economics of retrieval and summarization: instead of aggressively compressing everything before each call, a system designer can sometimes keep more of the raw evidence in view.

The sparse MoE design is the other half of the pitch. With 600 billion parameters in total but 27 billion active per token, Step 5 aims to offer the representational breadth of a giant model without paying dense-model inference costs on every token [1]. The ratio is central to why the release is competitive: dense 600B inference would be prohibitive for many teams, while a sparse model can concentrate compute on selected expert pathways. That does not make Step 5 cheap, but it changes the cost curve.

Open weights, not a hobbyist release

The phrase “open weights” needs precision here. StepFun has made the model weights publicly available, but this is not the same as a conventional open-source software release under a standard permissive license. The Hugging Face model card lists the license field as “other” and says the model is governed by the StepFun Model License 1.0, described as Apache-2.0-style with additional attribution and commercial conditions [1]. TSN Media emphasizes the same point: the release is not plain Apache 2.0 [5].

Those conditions matter for commercial users. TSN Media reports that large commercial products crossing 100 million monthly active users or more than $20 million in monthly revenue must display “Powered by StepFun AI,” while separate commercial licensing is required for certain hardware-device makers and model-as-a-service businesses above a revenue threshold [5]. AI Tamer also notes that the license treats models trained to a material extent on Step 5 outputs as derivative works, a significant clause for companies thinking about distillation [4].

So the practical conclusion is nuanced. Researchers, infrastructure teams and startups can get unusually large-scale model access outside a closed API, but the release is not a frictionless public-domain gift. Anyone building a product around Step 5 should read the license before fine-tuning, distilling, reselling access, or embedding the model into a device.

The infrastructure reality check

The self-hosting story is powerful, but the hardware requirements are substantial. StepFun’s deployment table lists vLLM FP8 serving on one B300 node with four GPUs or one H200 node with eight GPUs; BF16 serving is listed for one B300 node with eight GPUs or two H200 nodes with 16 GPUs [1]. MarkTechPost summarizes the same data-center reality, noting that FP8 needs 4x B300 or 8x H200, while BF16 needs 8x B300 or 16x H200 [2].

That means Step 5 Preview is “open” in the sense that the weights are obtainable, not in the sense that every developer can run it locally. The likely early users are AI infrastructure companies, large enterprises, research labs, agent-platform builders and cloud teams that already operate multi-GPU clusters. For them, the release can be strategically valuable: a private deployment can reduce data-governance concerns, allow lower-level optimization, and make model behavior easier to integrate into internal systems.

Even for those teams, the one-million-token context window should be treated as a capacity target rather than a free lunch. Long-context inference consumes memory through the key-value cache, and StepFun’s own deployment notes warn that actual memory requirements, usable context length and concurrency vary with checkpoint, inference framework and cache allocation [1]. Temperature2 makes the same point more bluntly: the listed layouts load the weights, but do not guarantee that a full one-million-token request will be practical once runtime buffers and KV cache are included [6].

Performance claims need independent replication

StepFun reports benchmark results across coding, agentic tasks, professional knowledge work and finance, but several recent reports stress that many headline charts are still vendor-reported [1][5]. MarkTechPost cites DeepSWE v1.1 at 67.7, Terminal-Bench 4.0 at 33.3, and an Artificial Analysis Intelligence Index score around 43.7 to 44 depending on the reference point [2]. Temperature2 also frames Step 5’s Artificial Analysis result as a cost-for-intelligence argument, not a claim that it leads the overall board [6].

That distinction is important. The release gives the community a path to test Step 5 Preview on the published weights rather than relying solely on StepFun’s hosted system or launch slides. For a model aimed at agentic work, the real evaluation will not be one benchmark table. It will be whether the model can survive messy repository edits, long tool loops, multi-document synthesis, finance-heavy reasoning, and private workloads where evaluation data is not neatly packaged.

The most interesting external tests will likely focus on three questions. First, do the public weights reproduce the reported coding and finance numbers? Second, how much useful context can teams actually serve at acceptable latency and cost? Third, does the 27B-active architecture deliver a meaningful cost advantage over dense or heavier-active open alternatives in real agent deployments?

Why this release matters

StepFun’s move adds pressure to a market where the largest and most capable models are often gated behind closed APIs. A 600B open-weight model with a one-million-token context window gives developers more room to build custom agents without surrendering all control over infrastructure, data flow or model adaptation [2][3]. It also intensifies competition among Chinese and global labs racing to offer large, long-context, agent-oriented models.

Still, the release should be understood as a data-center event, not a consumer-device event. The model’s active-parameter ratio may reduce inference cost relative to a similarly sized dense model, but the checkpoints, GPU requirements and context memory still put it in enterprise territory [1][2]. The license adds another layer of strategic calculation: open access, yes; unrestricted commercial reuse, no [5].

The Step 5 story is therefore bigger than one model card. It is a sign that frontier-scale access is becoming more plural: not just one hosted API, not just one proprietary assistant, but a growing menu of downloadable systems that sophisticated teams can tune, serve and benchmark themselves. If Step 5 Preview proves stable outside StepFun’s own environment, its one-million-token window and 27B-active MoE design could become a serious building block for customizable long-horizon agents.

Sources from the last 72 hours

1. [1]stepfun-ai/Step-5-Preview · Hugging FaceOct 15, 2026, 3:28 AM
2. [2]StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context - MarkTechPostOct 16, 2026, 2:00 AM
3. [3]StepFun releases Step 5 Preview open weights | Post-CutoffOct 15, 2026, 2:00 AM
4. [4]StepFun's Step 5 Preview weights are public, under its own license · AI TamerOct 15, 2026, 4:37 PM
5. [5]StepFun Publishes Open Weights for Step 5 Preview Under Its Own Licence, Not Apache 2.0Oct 15, 2026, 2:00 AM
6. [6]StepFun publishes 600B Step 5 Preview open weightsOct 15, 2026, 2:00 AM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.