Step-3.7-Flash: official model and deployment card

What is included in Step 3.7 Flash?

The publisher describes about 198B parameters in total, combining a 196B language backbone and a 1.8B vision encoder, with roughly 11B active parameters per token. The advertised context is 256K tokens and reasoning has low, medium and high levels. These architecture labels do not establish a local memory budget.

Source: stepfun-ai · pinned document ↗

Is the advertised throughput a desktop benchmark?

The introduction advertises up to 400 tokens per second, but does not identify a matching local workstation configuration in that sentence. We do not reuse it as a measured Spark, Apple or consumer-GPU result. Its capability scores also name distinct tasks and tool settings, such as SimpleVQA with search and V* with Python.

Source: stepfun-ai · pinned document ↗

Which local engine paths are documented?

The card separates vLLM, SGLang, Transformers and llama.cpp paths. Transformers is presented as a debugging and verification route. The vLLM recipe has explicit precision and multi-GPU choices; the existence of a llama.cpp section does not mean a generic GGUF or an arbitrary runtime revision supports the model.

Source: stepfun-ai · pinned document ↗

What must be fixed before testing?

Choose the exact weight format, reasoning level, parser and engine version, then retain prompt/output lengths and cache policy. A vision request also needs its image preprocessing and encoder placement. Keep a short load check separate from a timed or quality evaluation, and attribute all figures here to the publisher until an independent run exists.

Source: stepfun-ai · pinned document ↗

What does the source specify?

These declarations belong to the cited document. A planning minimum, model limit and measured runtime allocation are different quantities.