AI daily roundup: OpenCode rolls DeepSeek V4.1 Flash to build 1017
A daily roundup says OpenCode updated DeepSeek V4.1 Flash to build 1017 on its Go and Free plans with no user action needed, and notes Qoder's free window through Oct 24.
AI Daily · Oct 20 ⚙️ DEV 🔹 OpenCode updated DeepSeek V4.1 Flash to build 1017 across Go and Free plans; no user action needed. 🔹 Qoder is offering DeepSeek V4.1-Flash-1017 free for a limited time through Oct 24. 🔹 Command Code launched MiniMax M3.1 Flash, free for all subscription tiers. 🔹 Notion AI adds Auto mode to cut task costs, a new credits system, and MCP memory sync. 🧠 MODELS 🔹 MiniMax released M3.1 Flash: stronger agent, coding and computer-use tasks at 70% lower agent-task cost; weights later, Pro coming. 🔹 Ant Group open-sourced 5 Ming-3.0 speech models, from a 0.5B speech-generation model to a 126B audio-understanding model. 🔹 Inception released diffusion language model Mercury 3, claiming wins over GPT-6 Luna and Claude Haiku 5.5 at 4-8x speed. 🔹 Meta published a tech report for Muse Glimmer, a 30B open model targeting agent tasks on consumer hardware. 🏢 INDUSTRY 🔹 Meta completed phase 1 of gigawatt cluster Prometheus, tens of thousands of GPUs training Muse, scaling toward ~1M GPUs. 🔹 Microsoft extended Project Perception to police internal enterprise agents, continuously assessing risk and adjusting permissions. 🔹 Doubao PR chief disputed claims that most of China's compute makes short-drama video, saying ByteDance spends ~10% on video models. ⚠️ US DHS ordered federal civilian agencies to stop procuring DeepSeek and remove all its products, including third-party tools built on it.