Skip to content
Research and DevelopmentResearch and Development

The watch: week of 26 October

We follow a fixed list of sources: NVIDIA, CUDA, Ollama, OpenVINO, PyTorch, ComfyUI, llama.cpp, NeMo Relay, PAIR and Claude Code. This week turned up 101 new items. Most were noise; a few matter.

What we acted on

Section titled “What we acted on”
  • Claude Code 2.1.304 to 2.1.310. Two permission fixes. In 2.1.309, an Edit deny rule on a path also blocks a shell redirect to that path in commands that hold a subshell, where it didn’t before. And on Windows, pre-tool hooks no longer let a call through when Git Bash is missing: 2.1.309 fixes this for hooks set to run in PowerShell, and 2.1.310 for hooks whose command is only quoted strings. Our agent harnesses rely on deny rules and hooks, so we’ll test our deny rules once our install moves to these versions. Both changes can only make the checks stricter. See the Claude Code releases.
  • NeMo Relay 0.10.3 and 0.10.4. The tracing tool’s stable releases add hardening and a hook fix. Our notes cover 0.10.0 to 0.10.2, so we’ll add these two. The 0.11.0 release candidates are pre-release and contain breaking changes, so we are waiting for the stable build. See NeMo Relay.

Worth knowing

Section titled “Worth knowing”
  • Ollama 0.40.7 to 0.40.11: updated llama.cpp, speculative-decoding draft support and clearer streaming errors. No changes to scheduling, memory or CUDA support. Releases.
  • Nemotron 3.5 Bolt E4B: a small vision-language model with tool use; licence still being reconciled before any trial.
  • d3: classification-style decision models that could suit routing questions; licence unconfirmed.
  • Flint: a mixture-of-experts coding model under Apache-2.0; its benchmark figure is the vendor’s own.
  • MotionBricks: NVIDIA motion generation for characters; weights, code and sample data carry different licences, and the samples are evaluation-only.
  • Docker Model Runner 1.2.9: an alternative llama.cpp runner; reference only.
  • NVIDIA skills catalogue: new and changed skills, including Run:ai, are being synced into our catalogue.

Skipped this week: about 45 items across NVIDIA’s blog, model hub, skills metadata, llama.cpp builds and Docker news.

From our own bench this week

Section titled “From our own bench this week”

The week’s biggest finding was in our own measurements. Three short runs on a rented T4 showed that interlock’s generate ladder, in 1.0.5 and earlier, timed a prompt cache instead of the prompt, and counted queued requests as parallel ones. It’s fixed in interlock 1.0.5 to 1.0.7, and issue 31 tells the whole story. The newsletter also started publishing itself to Hugging Face, npm and PyPI.

How the watch works

Section titled “How the watch works”

A scheduled script lists what is new on a versioned set of sources; a scout reads the primary notes and ranks each item. Nothing is installed or adopted from the watch alone; licences come from each model’s own pages.