PulseAugur / Brief
EN
LIVE 19:11:34

Brief

last 24h
[3/3] 222 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

  1. TOOL · 36氪 (36Kr) 中文(ZH) ·

    OceanGrid Information Completes Step 3.7 Flash Adaptation for Step Star

    Jieyue Xingchen has released its Step 3.7 Flash model, which has been fully adapted and optimized by the Haiguang team. This integration allows developers to efficiently run the model on Haiguang's platform, facilitating the creation of multimodal agents and other AI applications. The collaboration aims to reduce costs associated with accessing and orchestrating AI models. AI

    IMPACT Enables efficient deployment of advanced AI models, potentially lowering barriers for AI application development.

  2. TOOL · Mastodon — fosstodon.org Deutsch(DE) ·

    RT @vllm_project: Congratulations to @StepFunai on the release of Step-3.7-Flash with day-0 support in vLLM. - 198B sparses MoE-Vision-Lang

    StepFunai has released Step-3.7-Flash, a 198 billion parameter sparse Mixture-of-Experts model. This new vision-language model offers day-zero support within the vLLM inference engine. The integration with vLLM is highlighted as a key feature for efficient deployment. AI

    IMPACT Enables efficient deployment of a large sparse MoE vision-language model.

  3. SIGNIFICANT · Hugging Face Trending Models Română(RO) · · [4 sources]

    stepfun-ai/Step-3.7-Flash

    JieYue (StepFun) has released and open-sourced Step 3.7 Flash, a new generation of Flash models optimized for agentic tasks, coding, and multimodal workflows. This Mixture-of-Experts (MoE) model features 196 billion total parameters with only 11 billion active, and includes a built-in vision transformer. Benchmarks show it performing competitively with models like DeepSeek V4 Flash and Gemini 3.5 Flash on coding tasks, and close to GPT-4.5 on a QA benchmark, making it a strong option for local agent applications. AI

    IMPACT Sets a new benchmark for flash-tier models in agentic and coding tasks, potentially lowering costs for multimodal AI applications.