DeepSeek’s Cache-First Play Transforms LLM Costs

Stepfun (阶跃星辰) has open sourced Step 3.7 Flash, a new Flash-series model optimized for production agent workloads with a sparse MoE architecture. The model totals 196B parameters plus a 1.8B ViT and 11B activation parameters, and claims up to 400 tokens/sec generation—targeted for high-frequency, multi-turn, low-latency agent applications. Key features include native multimodal understanding and execution (UI, charts, documents, images), strengthened web and visual search for cross-source evidence retrieval, robust tool calling and orchestration across long multi-step workflows, and compatibility optimizations for major agent frameworks and tool protocols (e.g., Claude Code, KiloCode, RooCode, MCP/Skills). Code and model artifacts are available on GitHub, Hugging Face, Modelscope, and Stepfun’s platform.

StepFun released Step 3.7 Flash, a multimodal mixture-of-experts (MoE) model with 196 billion total parameters but only 11 billion active, designed to run locally on a system with 128 GB of RAM. The model includes a built-in 1.8B ViT visual encoder and targets the “flash” class of efficient large models. Reported benchmarks show competitive performance: SWE-Bench Pro 56.26% (slightly ahead of DeepSeek V4 Flash and similar to Gemini 3.5 Flash) and a DeepSearchQA F1 of 92.82%, positioning it as a strong local inference option. This matters for developers and organizations seeking high-capability multimodal models that can run on modest hardware without cloud dependency. Key players: StepFun; competing models referenced include DeepSeek and Gemini.