Stepfun AI has launched Step 3.7 Flash, a 198B sparse Mixture-of-Experts vision-language model optimized for high-speed agentic coding and browser automation.

// 1h agoMODEL RELEASE

Stepfun AI has launched Step 3.7 Flash, a 198B sparse Mixture-of-Experts vision-language model optimized for high-speed agentic coding and browser automation.

Stepfun AI has released Step 3.7 Flash, a highly efficient 198B-parameter sparse Mixture-of-Experts (MoE) vision-language model designed specifically for real-world agentic workflows like browser automation and coding. Activating just 11B parameters per token to achieve blazing-fast speeds of up to 400 tokens per second, the model supports a massive 256K context window and native multimodal capabilities, allowing it to process text, GUIs, and wireframes. Released under the Apache 2.0 license, it is available for local deployment and via platforms like OpenRouter, featuring unique selectable reasoning levels (low, medium, high) that give developers granular control over speed, cost, and analytical depth.

// ANALYSIS

Offering selectable reasoning levels in an open-weight, high-speed MoE vision-language model is a major breakthrough for agentic applications, enabling developers to build cost-effective architectures where deep thinking is dynamically engaged only when needed.

* Extreme high throughput of up to 400 tokens per second is a perfect match for real-time agentic tool use and iterative browser tasks.

* Three selectable reasoning modes (low, medium, high) allow developers to customize execution speed and cost to match the specific complexity of a subtask.

* Open-weight release under Apache 2.0 invites wide integration, though the 198B total parameter count still requires high-end hardware with 128GB+ unified memory for local runs.

* Native multimodal and GUI design makes it exceptionally capable of parsing complex wireframes, UI layouts, and charts without requiring extra visual-indexing pipelines.

// TAGS

step-3.7-flashstepfunmoevision-languageopen-sourceagentmachine-learningdeep-learning

DISCOVERED

1h ago

2026-06-01

PUBLISHED

1h ago

2026-06-01

RELEVANCE

9/ 10

AUTHOR

Bijan Bowen

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

SECURITY21m ago

A pseudonymous security researcher has released six Windows zero-day exploits in a retaliatory campaign against Microsoft.

A pseudonymous security researcher has publicly released six Windows zero-day exploits, including BlueHammer and RedSun, in a retaliatory campaign against the Microsoft Security Response Center. This controversial disclosure has led to immediate platform bans for the researcher, active real-world exploitation of the vulnerabilities, and a public call from Microsoft reminding the community about the importance of coordinated vulnerability disclosure.

NEWS33m ago

Liran Tal flags agent skill security risks

At the AI Native Dev conference in London, security researcher Liran Tal warned that third-party AI "Agent Skills" pose significant supply-chain risks if installed without review. Because these skills control how AI agents interact with local environments and tools, compromised skills could enable severe security breaches and data exfiltration.

NEWS33m ago

MiniMax M3's SWE-bench superiority claims fall short in real-world production stress tests.

A hands-on developer test of MiniMax M3—currently integrated for free on OpenCode with claims of surpassing GPT-5.5 on SWE-bench—reveals significant instability and failure under real production workloads. During real-world execution, the model broke a push-to-talk feature, glitched a game through the floor, and failed to correctly render video after multiple attempts.

Stepfun AI has launched Step 3.7 Flash, a 198B sparse Mixture-of-Experts vision-language model optimized for high-speed agentic coding and browser automation.