• ’ve been running local LLMs for a while now. If you’re anything like me, you know the feeling — you download a new model, hit “load,” and watch your machine choke because it doesn’t have enough RAM. It’s frustrating, especially when you know you have a more powerful machine sitting somewhere else.

  That’s exactly the situation I was in. And then I got early access to LM Studio Link.

• I have two Apple Silicon machines at home:

  • Mac Mini M4–16GB RAM. My daily driver. Compact, quiet, always on.
  • MacBook Pro M4 Max — 64GB RAM. The beast. But it’s not always at my desk.

  The Mini handles everyday tasks beautifully. But when it comes to running serious models like OpenAI’s GPT-OSS 20B, Qwen 3.5 35B, or Llama 3 70B? Not a chance. These models need far more memory than 16GB can offer.

• What Exactly Is LM Studio Link?

  Think of it as your own private AI network.

  LM Studio Link lets you connect multiple machines running LM Studio (or its headless counterpart llmster) into a secure mesh network. Once connected, you can load models on any remote device and use them from any other device — as if the model were running locally.

• he key details that matter:

  • End-to-end encrypted — built on top of Tailscale mesh VPNs using WireGuard protocol
  • No ports exposed — your machines never touch the public internet
  • Zero config — works behind firewalls, NAT, and corporate networks without manual port forwarding
  • Identity-based access — no API keys to manage or rotate
  • Runs entirely in userspace — doesn’t modify any global system settings
• On the M4 Max (the host), I opened LM Studio, logged in to my account, and enabled Link. The machine appeared on my network within seconds. I clicked the Link button at the bottom, clicked the "Create your link " button, connected to my Google Account, selected "GUI is present on the machine," and followed the instructions on the other machine, in my case, a Mac Mini!

• On the Mac Mini (the client): I logged in with the same account. Under “Network Devices,” my M4 Max appeared as a connected device. I could see all the models available on it — GPT-OSS 20B ,Glm 4.7 Flash, Nemotron 3 Nano, Qwen3 Coder 30B, Gemma 3 27B Instruct, and a dozen more.

  Press enter or click to view image in full size

• That’s when it hit me. I was running a 20-billion parameter model on a machine with 16GB of RAM. The chats were stored locally on my Mini. The inference was happening on my M4 Max. And the entire thing was encrypted end-to-end.

  No cloud provider involved. No tokens are being counted against a billing meter. No data leaving my local network.
• You stop buying duplicate hardware. Instead of maxing out every machine, you invest in one powerhouse and share its compute across your devices. My Mac Mini doesn’t need 64GB of RAM anymore.
• Privacy stays intact. This isn’t a cloud relay. Your prompts and model weights travel directly between your devices over encrypted peer-to-peer connections. Neither Tailscale nor LM Studio’s servers can see your data. Only device metadata for connection setup touches their backend.

• The home office setup: Keep a GPU rig or high-memory Mac in one room, use lightweight laptops everywhere else. Your inference happens on the heavy machine regardless of where you’re sitting.

  Team scenarios: LM Studio Link is free for up to 2 users with 5 devices each — a total of 10 devices. A small team could share a single powerful inference server without any cloud costs.
• Edge + power hybrid: Use a Raspberry Pi or a lightweight device as a thin client for AI tasks, with the heavy lifting handled on a remote workstation. I’m already thinking about trying this with my OpenClaw setup.

• Getting Started

  If you want to try LM Studio Link:

  1. Download LM Studio from lmstudio.ai
  2. Request access to Link at lmstudio.ai/link
  3. Install LM Studio on both your host (powerful machine) and client (lightweight machine)

10 more annotations...

• We have identified industrial-scale campaigns by three AI laboratories—DeepSeek, Moonshot, and MiniMax—to illicitly extract Claude’s capabilities to improve their own models. These labs generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts, in violation of our terms of service and regional access restrictions.
• These labs used a technique called “distillation,” which involves training a less capable model on the outputs of a stronger one. Distillation is a widely used and legitimate training method. For example, frontier AI labs routinely distill their own models to create smaller, cheaper versions for their customers. But distillation can also be used for illicit purposes: competitors can use it to acquire powerful capabilities from other labs in a fraction of the time, and at a fraction of the cost, that it would take to develop them independently.
• Illicitly distilled models lack necessary safeguards, creating significant national security risks. Anthropic and other US companies build systems that prevent state and non-state actors from using AI to, for example, develop bioweapons or carry out malicious cyber activities. Models built through illicit distillation are unlikely to retain those safeguards,
• The three distillation campaigns detailed below followed a similar playbook, using fraudulent accounts and proxy services to access Claude at scale while evading detection

• Each campaign targeted Claude's most differentiated capabilities: agentic reasoning, tool use, and coding.

  DeepSeek

  Scale: Over 150,000 exchanges

  The operation targeted:

  • Reasoning capabilities across diverse tasks
  • Rubric-based grading tasks that made Claude function as a reward model for reinforcement learning
  • Creating censorship-safe alternatives to policy sensitive queries

• How distillers access frontier models

  For national security reasons, Anthropic does not currently offer commercial access to Claude in China, or to subsidiaries of their companies located outside of the country.

  To circumvent this, labs use commercial proxy services which resell access to Claude and other frontier AI models at scale.

• A prompt like the following (which approximates similar prompts we have seen used repetitively and at scale) may seem benign on its own:

  You are an expert data analyst combining statistical rigor with deep domain knowledge. Your goal is to deliver data-driven insights — not summaries or visualizations — grounded in real data and supported by complete and transparent reasoning.

  But when variations of that prompt arrive tens of thousands of times across hundreds of coordinated accounts, all targeting the same narrow capability, the pattern becomes clear.
• Detection. We have built several classifiers and behavioral fingerprinting systems designed to identify distillation attack patterns in API traffic.

6 more annotations...