According to DeepSeek on X, the company open sourced its in-house inference framework DeepInfer, which powers DeepSeek-V4.1-Flash at scale to cut compute costs and latency.

According to AI News on X, DeepSeek V4.1 Flash achieved 77% on the DeepSWE coding benchmark at under $0.20 per task, signaling cost efficient software automation.

According to DeepSeek on X, DeepSeek-V4.1-Flash-1017 delivers stronger benchmark performance, faster task completion, and shorter outputs, and is live on its website, app, and API with optimizations for DeepSeek Harness v0.2.1-rc.1, Tencent CodeBuddy, and WorkBuddy.

According to DeepLearningAI, DeepSeek Flash cuts cache to 890B per token, scores 39 on AA Index, and lowers cost to $0.27 per task versus $0.67.

According to DeepLearning.AI, GLM-5.3 reached 12% vulnerability exploits vs Claude Mythos at 14%, plus updates on Xiaomi leader, Gemini 3.8 Live, and more.

According to DeepSeekHarness, packaged desktop apps for macOS and Windows are live, with Linux via npm @ deepseek-ai/dsh, enabling local AI workflows.

According to DeepLearningAI, OpenAI ran 10,000 agents on Navier–Stokes and DeepSeek launched V4.1 Flash for cheaper long-context workloads.

According to DeepLearningAI, sloppy sandboxing, not runaway agents, drove incidents; plus disputes on OpenAI math proof and Anthropic IP claims.

According to @galnagli, Anthropic reveals how actors probed Claude for cyber, influence, and bio misuse, and how safeguards disrupted all cases.

According to TheRundownAI, Anthropic revealed global Claude misuse, DeepSeek cut prices, and guides covered ChatGPT Work and Google Astra uses.

According to @CNBC, Anthropic disrupted China-based labs allegedly using Claude for illicit distillation, naming Alibaba, Moonshot, and DeepSeek.

According to @CNBC, Anthropic alleges Chinese labs trained on millions of Claude chats, raising IP risks and compliance gaps for model builders.

According to TheRundownAI, DeepSeek V4.1 Flash tops GPT 5.6 Sol and Claude Opus 5 on agentic, coding, cyber tests with ultra low $0.15M in, $0.60M out.

According to KyeGomezB, DeepSeek’s causal encoder‑decoder MoE cuts active params to 8B in, 16B out, promising lower cost and top benchmarks.

According to @deepseek_ai, V4.1-Flash adds native vision, faster inference, and higher throughput, scaling to larger models for real-world deployment.

According to @_avichawla, open-weight models incur higher real costs per solved task than pricing suggests across 16 benchmarks, per Martian’s AI Frontier.

According to DeepLearning.AI, DeepSeek’s open source harness logs tool calls, prompts, and subagents, letting developers reproduce and benchmark LLM performance.

According to DeepLearningAI, this week covers GLM 5.3 security, real time throughput gains, DeepSeek V4 Pro tools, and Self GC for long contexts.

According to @CNBC, DeepSeek founder Liang Wenfeng expands a quant AI empire to navigate China’s choppy IPO market, focusing on data, models, and deal flow.

According to @godofprompt, OpenAI cut GPT5.6 Luna price 80% while DeepSeek shipped a stronger model; tests show DeepSeek wins on canvas tasks.