Urgent.News

What's breaking now, across hundreds of outlets.

Tech

Open-Source Models Are Ready for Enterprise…Here's How to Deploy Them Safely

The key to safe deployment lies in self-hosting open models, controlling data transmission, and ensuring model provenance, patch cadence, and access controls.

Open-Source Models Are Ready for Enterprise…Here's How to Deploy Them Safely

A year ago, talking about DeepSeek or Qwen in an enterprise meeting would have met with silence. Now, those two models account for nearly 15% of the global AI market, up from just 1% at the beginning of 2025. The question has shifted from "Are these models good enough?" to "How can we deploy them without creating a liability?" Enterprises are more focused on the cost, with teams reporting an 80% reduction in inference bills compared to using only premium models. However, this cost advantage is contingent on proper deployment.

The skepticism surrounding open models is not unfounded. Early in 2025, Florida's state CFO banned DeepSeek, citing it as a national security risk, and several federal agencies followed suit. Security researchers discovered issues like weak encryption, potential SQL injection flaws, and undisclosed data transmissions in DeepSeek's R1 model.

The issue is twofold: data potentially reaching the Chinese government and geopolitical hostility towards anything China-linked. The distinction is crucial, as bans often target the consumer-facing DeepSeek app running on Chinese servers, not the open-weight model used in-house.

Organizations that have successfully deployed open models have self-hosted MIT- or Apache-licensed models, ensuring no prompt data leaves their network. This approach eliminates most risks such as data transmission to foreign servers, opaque logging, and unclear retention. The remaining risk is a conventional vendor-risk conversation, focusing on model provenance, patch cadence, and access controls, similar to the due diligence applied to other software.

However, self-hosting has its own challenges. It doesn't eliminate the risk of a single point of failure, but it does improve security by atomizing compute, model execution, and data handling. This strategy, as demonstrated by Aphanarc, prevents any single component from holding the complete picture of sensitive information. The calculus for CTOs has evolved.

While some companies, like Coinbase and Cursor, have publicly adopted Chinese-origin models to cut costs, U.S. lawmakers are now scrutinizing the use of such models in American companies. The scrutiny is real, and the decisions enterprises make today will be the ones they must defend in the future.

Rewritten by urgent.news; may contain errors — check the original.

Read the original at hackernoon.com →

Report a problem with this story

More in Tech

The State of Context Engineering

This was interesting to me because I was under the naïve impression that context management was a thing of the past, but in actuality it's the opposite: context is here to stay, and managing it has…

writing about pratt parsing so I remember it

  • Pratt parsing extends shunting yard algorithm with call stack for subexpressions.
  • C code example demonstrates basic pratt-style parsing of numbers and operators.
  • Operator behavior defined in separate functions allows easy addition of new operators.

Kimi K3.1 Pro API: endpoints, tool calls, cache, and errors

The Kimi K3.1 Pro API runs at https://api.moonshot.ai/v1 with the model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. It speaks three transmission formats: OpenAI Chat Completions, Anthropic's…

  • Kimi K3.1 Pro API accessible at https://api.moonshot.ai/v1
  • API supports OpenAI, Anthropic, and custom response formats
  • Cache management and usage logging vary across API surfaces

How ATS Parsers Read Work Experience Dates, Career Gaps & Overlapping Roles (Greenhouse, Workday & iCIMS Breakdown)

When a software engineer or product manager with seven years of full-time experience gets auto-rejected within 90 seconds of applying for a role requiring "5+ years of experience," the culprit is…

  • ATS parsers require explicit start and end dates in ISO-8601 format (YYYY-MM)
  • Work experience calculation uses regex tokenizers for month and year pairs
  • Proper date formatting prevents undercounting in work experience totals

2026 Node.js Summarize Text API Chat Completions Example for Support Answers

Short answer: for a support agent answering from a private knowledge base, the difficult part is not producing a shorter paragraph.

  • Use Node.js to implement API with chat completions for support answers
  • Specify contract in prompt and validate JSON output fields
  • Set temperature 0.3-0.5, use GPT-4o-mini, and handle rate limits

More from Monday 26 October →