Step 5 Preview: Your Checklist Before the Oct 15 Open-Weight Release

Last updated: October 11, 2026

On 15 October 2026, four days from now, Shanghai-based StepFun plans to publish the full open weights of Step 5 Preview — its 600-billion-parameter flagship model for agentic work. The hosted API has been live since September, the model is already topping OpenRouter’s trending chart, and the launch coverage is all excitement about the specs. Nobody has published the practical part: what you should actually do to be ready. This is that checklist.

Quick answer: Step 5's open weights land 15 Oct 2026. Your 6-step checklist: the 600B/27B model explained, the multi-GPU hardware reality, and API ($1/M) vs local — which fits you.

What Step 5 Preview actually is

Step 5 Preview is a sparse mixture-of-experts (MoE) model: 600 billion parameters in total, with only about 27 billion active for any given token. That is roughly 4.5% of the weights doing the work per token, which is the whole design bet — the knowledge capacity of a giant model with inference costs closer to a small one. StepFun describes this as its “Pareto Frontier” philosophy: optimal intelligence per dollar rather than raw scale.

The official spec sheet, confirmed on the model’s OpenRouter listing (released there 8 October 2026), reads:

SpecificationDetail
Context window1 million tokens, up to 64,000 output tokens
Inputs / outputText, images, and video in; text out
Reasoning effortLow, medium, and high settings
FeaturesTool calling, JSON mode and JSON Schema, prompt caching, streaming

The model is built for agentic work — long-horizon, multi-step tasks that span large codebases and documents. That positioning shows up in the real traffic: the top public apps on OpenRouter currently using it are Nous Research’s Hermes Agent, StepFun’s own Step-Code, Kilo Code, Cline, and OpenClaw — coding and agent tools, exactly the lane it targets.

Some background helps. StepFun (formally Shanghai Jieyue Xingchen) was founded on 6 April 2023 by Jiang Daxin, who spent 16 years at Microsoft rising to corporate vice president, with co-founders Zhu Yibo (ex-ByteDance AI infrastructure) and Jiao Binxing. It is grouped with Zhipu AI, Moonshot AI, MiniMax, Baichuan and 01.AI as one of China’s “Six Little Tigers” (AI in China’s analysis). StepFun announced Step 5 Preview on 20 September 2026, skipping Step 4 entirely — the company said the jump was too significant for an incremental number. Earlier open releases include Step 3.5 Flash (February 2026, released under Apache 2.0) and Step 3.7 Flash (May 2026).

What the Oct 15 weights release changes (and what it doesn’t)

Open weights change who controls the model, not how much it costs to run. That distinction is the most important sentence in this guide.

Today, you can only use Step 5 Preview through StepFun’s API (via its Step Plan tiers or gateways like OpenRouter) at listed prices of $1.00 per million input tokens and $2.70 per million output tokens, with cache reads at $0.05 per million. After the weights release, anyone can download the full model and run it on their own hardware — no per-token provider charge, full ability to fine-tune, quantise, and inspect it.

What the release does not do:

The benchmark story, honestly told

There is exactly one fully independent number you should trust right now: 44 on the Artificial Analysis Intelligence Index, ranked 27th of 653 models tracked, at a weighted cost of $0.71 per completed task. That is a real third-party measurement (reported via GenZTech’s October analysis), and it is the basis of the “intelligence per dollar” story.

Everything else is vendor-published and unverified:

And the honest caveat no launch piece mentions: as of 11 October 2026, no SWE-bench Verified score exists from anyone for Step 5 Preview — GenZTech’s AI Coding Leaderboard keeps it in “verifying” status without a rank. The weights release is precisely what will let independent evaluators fix that.

The 6-step Oct 15 checklist

This is the part nobody else wrote. Work through it before Thursday.

Step 1: Decide your route — API now or local on Oct 15

If you need hosted inference today, the API route is $1.00/$2.70 per million tokens on OpenRouter (model slug stepfun/step-5-preview, released there 8 October 2026). Cache-heavy agent loops drop toward $0.05/M for cached reads. If you only need to inspect the weights and already have compute access, waiting avoids buying Step Plan or API tokens entirely — and note that Step Plan’s monthly credits expire at month end and do not roll forward (Omidsaffari’s pricing analysis), so do not pre-buy a big tier this week.

Step 2: Do the hardware math honestly

600B parameters in BF16 is ~1.2 TB of weights before KV cache. Full precision means a multi-GPU server (think 8× 80GB-class GPUs or NVMe offloading). Quantised builds (8-bit, 4-bit, or the community GGUF/exl2 formats) cut that dramatically — but they do not exist yet for this model. Anyone telling you a home PC will run the full weights on day one is selling you something.

Step 3: Watch the quant and runtime pipeline, not the weights alone

A weights release becomes usable through conversions and runners. Bookmark and watch: the llama.cpp, vLLM, and Ollama repos for quantisation and serving support; the Hugging Face listing when it lands; and the community quantisers who converted Step 3.5 Flash earlier this year. Day-one value goes to whoever has the runner ready, not whoever has the download fastest.

Step 4: Use the API as your “before” measurement

Run your real workload through the API this week — the same prompts, the same agent loops — and log the cost and the quality. When the local build arrives, you will have an apples-to-apples comparison instead of vibes. The model’s reasoning effort setting (low/medium/high) is the first knob to tune; test it on the cheap before committing GPU hours.

Step 5: Track the benchmarks that will matter, not the ones already claimed

Ignore the launch-table scores above until independent labs re-run them on the released weights. The two numbers to watch: a SWE-bench Verified score from an independent evaluator, and Artificial Analysis’s cost-per-task figure as more tasks complete. Vendor benchmarks get you hype; independent re-runs get you decisions.

Step 6: Plan your first day-one workload

Step 5 Preview’s verified lane is agentic coding and long-context document work — the 1M-token window plus tool calling is the combination to exploit. Candidates: multi-file refactors across a whole repo, long-document analysis with tool use (StepFun’s own demo coordinated 950 web fetches in one agent action — vendor claim, but it tells you what the tooling supports), and finance-domain agent workflows where the FrontierFinance claim (vendor-reported) suggests strength.

Step 5 vs Kimi K3: the cost-per-task decision

The fairest comparison available uses Artificial Analysis’s independent numbers, reported October 2026:

ModelAA Intelligence IndexInput price (per M tokens)Cost per completed task
Step 5 Preview44$1.00$0.71
Kimi K3 (Moonshot)44$3.00$2.15
Claude Opus 545$5.00$5.68

Same intelligence score as Kimi K3, roughly one-third the cost per task and one-third the input price (AI in China’s analysis). Note the caveat: on StepFun’s own Terminal-Bench table the two models diverge sharply, which is exactly why this table uses the independent index instead. If your workload is agentic coding specifically, wait for the independent re-runs in Step 5 above before switching.

What not to expect on Oct 15

Three expectations to keep in check. First, do not expect a licence announcement to change the value overnight — Apache 2.0 (like Step 3.5 Flash) would be the best outcome, but until StepFun publishes the licence text, treat commercial use as uncertain. Second, do not expect the quantised builds on day one — the pipeline in Step 3 needs the weights first. Third, do not expect the price war to end here: MiniMax, GLM and the rest of the Six Little Tigers ship weekly, and the “intelligence per dollar” crown changes hands fast. The checklist above is designed to survive all of that.

FAQ

Is Step 5 Preview free?

The API is not free — Step Plan lists paid tiers with no trial, and token calls carry the published $1.00/$2.70 rates. The October 15 open weights remove the provider usage charge, but you still need the GPUs, serving and engineering to run them.

What licence will the Oct 15 weights have?

Not yet specified. StepFun’s Step 3.5 Flash shipped under Apache 2.0, which is encouraging, but wait for the actual licence text before planning commercial use.

Can I run the 600B weights on my home PC?

Not at full precision — 1.2 TB in BF16 needs multi-GPU server hardware. Home-PC feasibility depends on the quantised builds (GGUF, exl2) that the community produces after the release.

After Oct 15, is the API still worth paying for?

Often yes. The API stays the zero-hardware option with 1M-token context, tool calling and prompt caching, and cache reads at $0.05/M make long agent loops cheap. Run your “before” measurement in Step 4, then decide with numbers.

Sources and methodology

Claims rest on the sources above — StepFun’s official spec via OpenRouter and StepFun’s launch announcements, plus independent reporting dated September–October 2026. Vendor-published benchmark claims are labelled as such; independent measurements (Artificial Analysis) are stated as independent. Desk-researched, not hands-on tested — the open weights release on 15 October 2026, which this guide prepares for, had not happened at publication time. Verification date: 11 October 2026.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!