# stepfun-ai/Step-3.5-Flash

Fast, Sharp & Reliable Agentic Intelligence

Repository: https://github.com/stepfun-ai/Step-3.5-Flash
Canonical: https://ross.abutalabs.com/products/step-35-flash
Homepage: https://static.stepfun.com/blog/step-3.5-flash/
Language: C++
License: Apache-2.0
License Family: permissive
Last push: 2026-04-03T12:50:32+00:00

## Health v2 (maintenance only)
Score: 49/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 75, release rhythm 35, longevity 15
- inputs: {"age_days": 214, "days_push": 152, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2070, forks 87 (observed 2026-08-28T04:06:10.342724+00:00)

## What it is
Step 3.5 Flash is an open-source 196B-parameter sparse Mixture-of-Experts foundation model (11B active per token) from StepFun, focused on fast reasoning and agentic capabilities. The repository provides model weights, deployment guidance, and cookbooks for running it locally and integrating it with agent platforms like OpenClaw, Claude Code, and Roo Code.

## Use cases
- run a fast open-source reasoning model locally
- build coding agents that score high on SWE-bench
- self-host an LLM for tool-calling and long-horizon agent tasks
- serve an MoE model with high token throughput
- integrate an open model into Claude Code or Roo Code workflows
- deep research and multi-step task automation with an open model

## When to choose
- you need an open-weights model with frontier-level reasoning and agentic tool use
- you want high generation throughput (100-350 tok/s) for real-time agent interaction
- you need strong software engineering performance (74.4% SWE-bench Verified) from a self-hosted model
- you want efficient long-context (256K) inference with modest active parameters

## When to avoid
- you lack GPU infrastructure for a 196B-parameter model
- you need a small model for edge or CPU-only deployment
- you need native multimodal (image/video) input, which requires the newer Step 3.7 Flash
- you only need a hosted API and don't want to manage inference yourself

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, agent-framework, machine-learning
- domain: large-language-models, artificial-intelligence
- platform: python
- tags: moe, open-weights, reasoning-model, agentic-llm, multi-token-prediction, swe-bench, coding-agent, ai-agents, gpu, linux, docker

## Member repositories
- stepfun-ai/Step-3.5-Flash (main) score 49

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:10.342724+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:56:44.491244+00:00, confidence not recorded.
  - readme: https://github.com/stepfun-ai/Step-3.5-Flash (fetched 2026-08-28T04:06:10.342724+00:00, sha 08c275e50020)
  - homepage: https://static.stepfun.com/blog/step-3.5-flash/ (fetched 2026-08-29T10:37:02.475789+00:00, sha 6d524d9e53e9)
  - site_page: https://platform.stepfun.com/docs/zh/llm/reasoning (fetched 2026-08-29T10:37:02.486650+00:00, sha 7d11493bc717)
- Data as of 2026-08-30T08:39:29.467469+00:00.