StepFun releases Step-3.7-Flash: 198B-parameter MoE model with 256K context at $0.20/M input tokens
StepFun has released Step-3.7-Flash, a 198B-parameter sparse Mixture-of-Experts vision-language model that activates 11B parameters per token and delivers up to 400 tokens per second. The model supports a 256K context window, three selectable reasoning levels, and is priced at $0.20 per million input tokens (cache miss) and $1.15 per million output tokens.
Step-3.7-Flash — Quick Specs
StepFun Releases Step-3.7-Flash: 198B-Parameter MoE Model
StepFun has released Step-3.7-Flash, a 198B-parameter sparse Mixture-of-Experts (MoE) vision-language model that activates approximately 11B parameters per token. The model combines a 196B-parameter language backbone with a 1.8B-parameter vision encoder and delivers throughput of up to 400 tokens per second.
Core Specifications
Step-3.7-Flash supports a 256K context window and offers three selectable reasoning levels (low, medium, and high) for developers to balance speed, cost, and cognitive depth. The model is designed for production workloads requiring perception, search, and reasoning across multi-step workflows.
Benchmark Performance
According to StepFun, Step-3.7-Flash achieves:
- SimpleVQA (Search): 79.2 (first place)
- *V (Python)**: 95.3
- ClawEval-1.1: 67.1 (significantly ahead of second place at 59.8)
- SWE-Bench PRO: 56.3 (second place)
- Toolathlon: 49.5
- HLE w. Tool: 48.1
- Terminal-Bench 2.1: 59.5
- GDPVal-AA: 45.8
The ClawEval-1.1 score demonstrates high resistance to adversarial traps during multi-turn orchestration, while the SWE-Bench PRO result shows capability in tracing multi-file repositories and generating functional patches.
Pricing
Step-3.7-Flash pricing structure:
- Input (cache miss): $0.20 per million tokens
- Input (cache hit): $0.04 per million tokens
- Output: $1.15 per million tokens
Availability and Deployment
The model is available through StepFun's Open Platform (platform.stepfun.ai for global access and platform.stepfun.com for China), OpenRouter, and NVIDIA NIM. StepFun is partnering with DeepInfra, Fireworks AI, and Modal for expanded availability.
For local deployment, Step-3.7-Flash supports vLLM, SGLang, Hugging Face Transformers, and llama.cpp. The model can run on NVIDIA DGX Station, AMD Ryzen AI Max+ 395-based systems, and Mac Studio/MacBook Pro devices with at least 128GB unified memory.
StepFun model support has been integrated into the NVIDIA Nemo ecosystem, including AutoModel, Megatron Core, and Megatron Bridge.
Technical Architecture
The sparse MoE architecture activates only 11B of the total 198B parameters per token, enabling the claimed 400 tokens per second throughput. The 1.8B-parameter vision encoder provides native image understanding for processing UI wireframes, application GUIs, and data charts.
What This Means
Step-3.7-Flash targets the growing segment of agentic workflows that require multi-modal understanding and tool orchestration at production scale. The sparse MoE architecture delivers a favorable compute-to-capability ratio compared to dense models of similar performance, though real-world latency and throughput will depend heavily on deployment infrastructure. The $0.20/M input token pricing positions it competitively against frontier models from larger providers, particularly for workloads that benefit from the 256K context window and cache hit pricing of $0.04/M tokens.
Related Articles
LG AI Research Releases K-EXAONE 2.0, a 750B-Parameter Open-Weight MoE Model with 262K Context
LG AI Research has released K-EXAONE 2.0, a 750-billion-parameter mixture-of-experts language model with 37B active parameters, a 262,144-token context window, and support for 10 languages. The model is open-weighted under Apache 2.0 and claims competitive results against Qwen3.5, GLM-5.1, and DeepSeek-V4 Pro on reasoning, coding, and long-context benchmarks.
OpenAI Halts Parts of Astra Model Development After It Hit 'Critical' Cybersecurity Threshold
OpenAI disclosed that its in-development Astra model showed cyberattack capabilities strong enough that it cannot rule out a 'Critical' risk classification. The company has paused related internal activity and added security controls under its Preparedness Framework.
Mistral's 3B-Parameter Shieldstral Matches 20B Safety Model on Text Benchmarks
Mistral's new Shieldstral, a 3-billion-parameter open-weight safety classifier, posts an 84.9% F1 score on text benchmarks—tying OpenAI's GPT-OSS-Safeguard-20B, a model roughly seven times larger. The model lets operators define safety rules at runtime using plain-language yes/no questions instead of fixed taxonomies.
Mistral AI Releases Shieldstral-1.0-3B, a 3B-Parameter Policy-Adaptive Safety Classifier
Mistral AI has released Shieldstral-1.0-3B, a compact open-weight safety classifier that evaluates text and images against natural-language policies specified at inference time. The 3B model runs on a single GPU and reports F1 scores competitive with or exceeding larger moderation models like LlamaGuard-4-12B and GPT-OSS-Safeguard-20B on multiple benchmarks.
Comments
Loading...