StepFun’s Step 5 Preview was available by API in September and as downloadable weights from 15 October, so an October launch-night event did not mark a new model release. A new OrcaRouter account of the event adds useful detail on the model’s costs, hardware demands and a recently revised benchmark score, all of which matter to anyone weighing whether to try it.
StepFun Watch analysis
What happened
OrcaRouter’s 29 October article describes a StepFun and Together AI event held in San Francisco on 22 October. It says the panel discussed how Step 5 performs in agent workflows and what it takes to serve the model. The article gives the API date as 20 September and the weights release date as 15 October, making the event a chance to discuss an existing model rather than a fresh launch.
The article says Artificial Analysis rebuilt its Intelligence Index on 27 and 28 October. Under the revised index, Step 5 Preview scores 29, compared with 44 under the previous version. Those scores use different test mixes, the article cautions, so the change is not evidence that the model itself got worse. It also says StepFun’s published agent benchmarks are vendor-run, and reports that the FP8 weights need roughly 773 GB of VRAM and at least four GPUs.
Why it matters
The dates and figures give builders a more practical picture than the launch-night label does. Step 5 is available to try, but self-hosting it calls for substantial hardware, while a score from a rebuilt index cannot be compared directly with the older result. The article also reports that StepFun’s Chinese-language cache-hit price doubled before the event, from ¥0.35 to ¥0.70 per million tokens. That detail could alter the economics for frequent users, though it is not a new change announced at the event.
Our read
The useful news here is the decision-making context, not a supposedly fresh release. OrcaRouter’s account is detailed, but several performance and cost figures are claims reported in that article, not independently established results. Builders should treat the revised benchmark as a new baseline and check current API pricing before doing their sums. The hardware requirement, meanwhile, makes “open weights” a rather less casual invitation than the phrase can suggest.
What to watch
- Whether independent, comparable evaluations of Step 5 Preview appear under the revised index.
- How API prices and cache discounts compare with the cost of running the weights.
- Whether StepFun or serving partners publish clearer evidence about reliability in long agent workflows.
Discussion spark: For an AI model with downloadable weights but a hardware footprint this large, is access meaningfully open, or does the API remain the only practical option for most builders?
Sources and evidence
not affiliated with or endorsed by StepFun