DeepSeek has released V4.1-Flash-1017, an open-weight model checkpoint that outperforms its predecessor on agentic task benchmarks while reducing per-task inference costs (The Decoder). The release is live across the DeepSeek website, app, and API. DeepSeek also opened new Search and Fetch APIs at $0.75 per 1,000 calls, expanding developer access to internal search infrastructure (The Decoder). The model demonstrates faster task completion and shorter outputs than the September V4.1-Flash checkpoint (MarkTechPost).
StepFun released Step 5 Preview as open-weight FP8 and BF16 checkpoints on Hugging Face and ModelScope (MarkTechPost). The sparse mixture-of-experts model contains 600B total parameters with 27B active per token and supports up to 1M token context windows, targeting long-horizon agentic applications. Perplexity open-sourced SPACE, its Firecracker microVM sandbox platform for AI agent task execution, under Apache 2.0 on GitHub (MarkTechPost), providing isolated compute environments with external credential and traffic control for sandbox-based agent workflows.