AI News — Morning — 2026-11-13

Larger models on local hardware

• Edge0 releases two large models for running locally — Downloadable models and a local runtime offer another option for coding work without sending prompts to a cloud provider.
  • Edge0 released 145-billion- and 120-billion-parameter models with its inference framework. The 145B model uses roughly 10 billion parameters per generation step and streams parts of the model between GPU memory, system memory and storage.
  • The 145B model card reports 98.3 generated tokens per second in its agent test on an RTX 5090 D with 24 GB of video memory. In a separate memory reading on that card, the model used 23.25 GiB of video memory and 38.27 GiB of system memory.
  • A lower-memory test constrained that same card to 16 GiB of video memory and 20 GiB of system memory. The card distinguishes those test budgets from minimum hardware requirements.
  • The model card documents NVIDIA builds for Linux and Windows; its Apple measurement does not establish that a macOS release is available. The framework code is Apache-licensed, while the model has separate terms.
  • Caveat: Edge0’s model license restricts commercial model services and independent commercial coding assistants without separate permission; speed and quality figures are the company’s own measurements. (HN 273 points · 77 comments · Edge0 announcement and model card)

Keeping coding work on spec

• A study asks coding assistants to explain deviations before writing code paper notes → — The technique gives a model a chance to notice where its planned shortcut conflicts with the actual requirements.

  • The researchers’ SatisFix method first asks an agent, a model running a multi-step task with tools, to report planned departures from the task and its reasons. It then asks the agent to reconsider each choice against the original description.
  • On the 113-task DeepSWE coding test, GPT-5.6-Luna passed 50.22% of tasks on average with SatisFix, versus 44.25% without it, over four runs. The researchers report an average cost of $0.897 per task versus $0.644 for the base agent.
  • With Qwen3.8-27B on the same test, the reported pass rate rose from 42.26% to 47.79%. The study also tested a separate coding task set and a second agent framework.
  • Caveat: These are benchmark results on selected models and tasks, not evidence that the prompt prevents missed requirements in every real project. (arXiv 2611.16289 · full paper)

• Coding assistants often keep working after their first passing fix paper notes → — Tracking intermediate patches shows the cost of extra work and the risk of breaking a passing fix.

  • Researchers saved each distinct repository patch and tested it offline against the coding benchmark’s hidden tests. In their main settings, a median 35–60% of steps in successful runs came after the first passing patch.
  • Much of that later work was verification or exploration. In 1–8% of runs that reached a passing patch, a later edit broke it and the final submission failed, depending on the setting.
  • The study covers five models, two agent frameworks and three coding task sets. Its main analysis focuses on Qwen3-Coder and Laguna-XS.2 on two SWE-bench datasets.
  • Caveat: The agents could not see the hidden tests that identified the first passing patch, so the paper does not provide a reliable real-time stop rule or justify skipping verification. (arXiv 2611.14960 · full paper)

Running repeatable agent work

• AgentRun adds a self-hosted console for repeatable agent jobs — It combines run recovery, spending limits and reusable notes around existing coding assistants.
  • The Apache-licensed project supports Claude Code, Codex, OpenCode and pi. Its local quickstart requires Node 24.22.3 or newer on macOS or Linux and supports Anthropic or OpenAI keys.
  • With learning enabled, the agent proposes notes, skills or workflows after a run; the software checks and grades candidates before saving them for later runs. The README describes approvals and spending caps as well.
  • In a company case study, its reviewer caught 42% of an answer key of 78 bugs in 30 held-out pull requests, versus 26% for the comparison product. The answer key pooled findings and used a model to grade them.
  • Caveat: The review comparison is the vendor’s small, model-graded test on its own repository; its result should not be assumed to carry over to another codebase. (HN 109 points · 29 comments · project README and case study)

Service reliability

• Claude reports an authentication outage and says it has recovered — Anyone who hit errors in Claude Code this morning can distinguish a provider incident from a local setup problem.
  • Anthropic began investigating authentication problems at 09:58 UTC and reported a deployed fix at 11:05 UTC on November 13.
  • Its incident page lists Claude Code, Claude Console, claude.ai and Claude Cowork as affected. At 11:58 UTC, the resolved update specifically said the issue affecting claude.ai and Cowork had been resolved.
  • Caveat: The final update names only claude.ai and Cowork, although the incident is marked resolved overall; check the status page for later changes. (HN 224 points · 163 comments · Anthropic status page)