DHS Orders Federal Agencies to Remove DeepSeek Products
DHS ordered federal agencies to remove DeepSeek products from their systems.
AI Weekly Recap
42 articles · 6 categories
Weekly pattern report
6 shifts that shaped AI this week
2026-10-17 → 2026-10-23
2026-W43 · 42 articles reviewed
The week in signals
- US bans DeepSeek from federal agencies; DeepSeek reportedly raises ~$15B from Tencent and CATL.
- DeepSeek open-sources DeepInfer, the engine behind its production inference.
- DeepSeek V4.1-Flash-1017 ships MIT-licensed, with search and fetch APIs at $0.75 per 1,000 calls.
- vLLM adds TPU, Humming kernels and elastic expert parallelism; llm-d reaches 1.0.
- OpenAI, Anthropic and Vercel all ship agent security or durability tooling.
- OpenAI and Claude models both land on Google Cloud.
The US banned DeepSeek from federal systems, and DeepSeek still reportedly raised about $15 billion. Its Oct 21 release of the DeepInfer inference engine and a cheaper, agent-tuned V4.1-Flash show an open-weight lab competing on serving cost, not only benchmarks.
The same pattern ran through the stack: vLLM, llm-d and Modal shipped serving and sandbox upgrades, while Vercel and Databricks shipped durable harnesses and agent memory. Security moved from policy to tooling, with OpenAI's containment write-up, Patch the Planet and Anthropic's OSS Scanner.
For platform teams, the durable lesson is that agent cost, safety and reliability now live in the runtime layer, and more of that layer is open source.
Prefer to watch?
LLM Digest Weekly Explainer: Oct 17-Oct 23, 2026
DeepSeek Ban Meets a $15B Raise 6 items
Washington cut DeepSeek out of federal systems the same week investors, including Tencent and CATL, reportedly backed it with about $15 billion.
US bans federal agencies from buying DeepSeek AI products
The ban also bars federal agencies from buying DeepSeek AI products.
Deepseek raises $15 bln in Tencent, CATL-backed funding round - Bloomberg
Bloomberg reports DeepSeek raised $15 billion in a round backed by Tencent and CATL.
DeepSeek doubles round to $15 billion, led by CATL and Tencent
The round doubled in size to $15 billion, led by CATL and Tencent.
China defends DeepSeek after US ban of its AI from government systems
China defended DeepSeek after the US removed its AI from government systems.
China's DeepSeek, MiniMax win approval for AI global use, state media-affiliated entity says
DeepSeek and MiniMax won approval for global use, per a state media-affiliated entity.
Open-Weight Models Move Up the Stack 7 items
Open-weight labs are shipping cheaper agent-tuned models and bundling the tools around them, not just weights.
DeepSeek Makes V4.1-Flash-1017 Official With MIT-Licensed Weights and Stronger Agentic Coding
DeepSeek made V4.1-Flash-1017 official with MIT-licensed weights and stronger agentic coding.
DeepSeek’s new AI model talks less, so a coding task costs 20 cents
The new model talks less, which brings a coding task down to about 20 cents.
DeepSeek prices web search for AI agents at $0.75 per 1,000 calls
DeepSeek prices its agent web search at $0.75 per 1,000 calls.
DeepSeek moves up the stack: V4.1-Flash update ships alongside its own Search and Fetch APIs for agents
V4.1-Flash ships alongside DeepSeek's own Search and Fetch APIs for agents.
Introducing Kimi K3.1 Pro
Kimi K3.1 Pro is on the Kimi API now; Moonshot promises full weights by Nov 4, 2026.
EmbeddingGemma 2 Expands beyond Text to Support Code, Images, Video, and Audio
Google's Apache 2.0 EmbeddingGemma 2 adds code, image, video and audio embeddings for RAG.
Latest open artifacts (#25): DeepSeek V4.1 Flash 1017, American-made models, and new regulatory controls in the US & China
Interconnects' roundup covers DeepSeek V4.1 Flash 1017, American-made open models and new US and China controls.
Inference Engines Go Open Source 9 items
The serving layer is being published: DeepSeek released its production engine while vLLM and llm-d shipped new backends and kernels.
DeepSeek Open-Sources DeepInfer, Its Production Inference Engine
DeepSeek open-sourced DeepInfer, the engine behind its production serving.
DeepSeek Open-Sources Framework Behind Its High-Margin AI Inference
The Information frames the release as the framework behind DeepSeek's high-margin inference.
GLM-5.3 Optimizations, Part 2: PCP, DCP, and Hybrid HiSparse on 16 H200s
vLLM serves GLM-5.3 on 16 H200s at 1M-token contexts, 1.7-3.2x per-user speed over data-parallel P/D on AgentX.
vLLM Delivers Day-0 Support for MiniMax-M3.1-Flash
vLLM added day-0 MiniMax-M3.1-Flash support on NVIDIA Blackwell, Rubin and AMD MI350.
Announcing Humming: Fast, Flexible Kernels for Quantized Inference
Ant Group and the vLLM community released Humming, a quantized inference kernel library.
Introducing vLLM TorchTPU
vllm-torchtpu is a new vLLM backend for Google Cloud TPUs with disaggregated serving.
Achieving near-zero-downtime scaling in Elastic Expert Parallelism
Asynchronous preparation and NIXL EP give vLLM near-zero-downtime Elastic Expert Parallelism scaling.
llm-d 1.0 Advances Open, Cost-Efficient AI Inference for Production Scale
llm-d 1.0 targets open, cost-efficient inference at production scale.
NVIDIA Vera CPUs double throughput for Modal Sandboxes
Modal benchmarked NVIDIA Vera CPUs on agentic Sandbox workloads and reports doubled throughput.
Agent Security Gets Operational 6 items
Labs and vendors are treating agents as both attack surface and vulnerability finders, with containment, scanning and runtime boundaries.
Securing our frontier model training environment
OpenAI describes stronger containment for frontier model training, including offline environments and high-compute red-teaming.
Patch the Planet: global response coordination as agents uncover critical vulnerabilities
OpenAI's Patch the Planet connects researchers, maintainers and responders on critical open-source vulnerabilities agents uncover.
Anthropic Adds Opt-In Security Scanning for Critical Open-Source Projects
Anthropic's free, opt-in OSS Scanner sends enrolled maintainers periodic security reports.
Scoop: Unvetted AI agents are accessing U.S. bank accounts
A scoop reports unvetted AI agents are accessing US bank accounts.
Red-Teaming OpenAI's Internal Agent Monitoring Systems
A write-up red-teams OpenAI's internal agent monitoring systems.
Nono: Agent runtime security. Kernel-enforced boundaries for AI agents
Nono enforces kernel-level boundaries on agent runtimes.
Production Agent Harnesses, Memory and Evals 9 items
Teams are hardening agents with durable execution, persistent memory and audited evaluators, while productivity evidence stays mixed.
Vercel Open-Sources durable-harness for Crash-Proof Coding Agents
Vercel open-sourced durable-harness for coding agents that survive crashes.
Vercel Labs Open-Sources fx: A Zig Coding Agent Designed to Be Embedded
Vercel Labs released fx, an experimental Zig coding agent that can be embedded, including as WebAssembly.
Introducing Agent Memory on Databricks
Databricks introduced agent memory on Agent Bricks.
MemCurrent: Designing Persistent Memory for Production Agents
Databricks' MemCurrent post covers designing persistent memory for production agents.
Context Engineering for AI Agents Starts with Retrieval
LangChain's Deep Agents harness uses retrieval techniques to keep agent context small and current.
How To Use Decision Models for Evals
Hamel Husain explains using decision models such as OpenAI's Decisions API for evals.
Who judges the judge? Three helpful ways to audit your LLM evaluators
Google Cloud gives three ways to audit LLM-as-a-judge evaluators.
Presentation: Actors for Agents: Patterns for Production AI
InfoQ talk applies the actor model to agent state ownership and message passing.
Study Finds AI-generated Code Fails to Boost Productivity Due to Downstream Bottlenecks
A Harvard study of 718 firms finds AI coding raises code volume without matching completed work.
Frontier Models Spread Across Clouds 5 items
Distribution is decoupling from lab: Google Cloud now offers OpenAI and Claude models, and agent skills moved to shared governance.
OpenAI models are coming to Google Cloud
OpenAI models and Codex workflows are coming to Gemini's Model Garden on Google Cloud.
Expanding Claude availability and deployment choice on Google Cloud
Google Cloud expands Claude availability and deployment options.
Claude Opus 5.5 and Claude Sonnet 5.5 are now available in Google Antigravity
Claude Opus 5.5 and Sonnet 5.5 are available in Google Antigravity.
Donating Agent Skills to the Agentic AI Foundation
Anthropic is donating Agent Skills to the Agentic AI Foundation.
How Commerce.com and Cloudflare are getting storefronts ready for AI agents
Commerce.com and Cloudflare enabled WebMCP shopping tools across nearly 100,000 domains.
The week, resolved into patterns