stepfun-ai/Step-3.5-Flash-Base · Hugging Face

Fields

id

sources/step-3-5-flash-hf-base

URL

https://huggingface.co/stepfun-ai/Step-3.5-Flash-Base

Kind

repo

Accessed

2026-10-19

notes

Read https://huggingface.co/stepfun-ai/Step-3.5-Flash-Base for the benchmark figures on models/step-3-5-flash. As recorded by the reader: Scores are for the released (post-trained) Step 3.5 Flash from the official release blog; the blog also confirms 74.4% SWE-bench Verified in prose. Terminal-Bench 2.0 (51.0) and HMMT/IMOAnswerBench/xbench results are reported but excluded (only Terminal-Bench 2.1 is on the allowed list). Base-model pretraining scores on the HF base card (e.g. MMLU 85.8, MMLU-Pro 62.3, GPQA 41.7) were not used because they describe the base checkpoint, not step-3.5-flash.

Notes

Read https://huggingface.co/stepfun-ai/Step-3.5-Flash-Base for the benchmark figures on models/step-3-5-flash. As recorded by the reader: Scores are for the released (post-trained) Step 3.5 Flash from the official release blog; the blog also confirms 74.4% SWE-bench Verified in prose. Terminal-Bench 2.0 (51.0) and HMMT/IMOAnswerBench/xbench results are reported but excluded (only Terminal-Bench 2.1 is on the allowed list). Base-model pretraining scores on the HF base card (e.g. MMLU 85.8, MMLU-Pro 62.3, GPQA 41.7) were not used because they describe the base checkpoint, not step-3.5-flash.

Cross-references

This record references no other records.