Model details
Step 3.5 Flash
Step 3.5 Flash is StepFun's open-weights foundation model released as a reasoning-focused system built around a sparse Mixture of Experts design. Rather than running all of its weight budget on every token, the architecture selectively activates only a fraction of the parameters at inference time, trading the compute cost of a much larger network for the responsiveness of a small one. OpenRouter's description of the model captures this intent directly, framing it as a reasoning model that remains speed-efficient even when prompts stretch into long-context territory, which suggests StepFun optimized the routing and inference path for sustained throughput on extended inputs.
For practitioners, the combination of open-weight availability, a large overall parameter pool, and a lean active footprint makes Step 3.5 Flash a practical fit for self-hosted deployments that need reasoning quality without paying the latency of a dense model at every step. Community interest in running it on compact hardware is already visible, with an NVIDIA DGX Spark forum thread demonstrating the model handling a 256k-context workload on a single device. Compared with denser open models of similar capability, the sparse design offers a route to long-context assistants, retrieval-heavy pipelines, and agent-style integrations where selective expert activation can keep response times low while still drawing on a deep underlying model.
Quick Info
Powered by- Provider
- Hugging Face
- Model key
- stepfun-ai/Step-3.5-Flash
- Release date
- Jan 29, 2026
- Last updated
- Feb 13, 2026
- Knowledge cutoff
- 2025-01
- Input modalities
- Output modalities
- Capabilities
Cost
- Input token cost
- $0.10
- Output token cost
- $0.30
Limits
- Output tokens
- 256,000 tokens
- Context window
- 262,144 tokens