AI engineering roundup: DeepSeek V4.1-Flash-1017 ships MIT weights, 77% DeepSWE
The October 17 roundup lists DeepSeek V4.1-Flash-1017 with MIT-licensed weights and a reported 77% on DeepSWE, among other items on TypeSafe, vLLM and OpenCode.
Today in AI Engineering (October 17) • DeepSeek releases V4.1-Flash-1017: MIT weights, reported 77% DeepSWE • TypeSafe releases 27B and 4B decider models with a new batching-oriented attention schema • vLLM reports 842 tokens/s for Kimi at batch 1 on GB300 NVL72 with DSpark • OpenCode reports Console breach affecting 1.4% of users, refunds underway • Infisical releases Agent Proxy: vault-backed credentials for CLI, MCP and model requests • Claude Code Desktop adds /btw side chats in separate tabs • Hugging Face Hub adds 150+ RL environments compatible with tools including Harbor, OpenEnv • Kitaru details cached tool-output replay without repeating state-changing calls • Theo posts Gemini cache-miss reproduction: 27% misses, costs roughly 50% above expectations