The watch: week of 13 October

Every Monday the studio checks twelve sources for anything new:

• NVIDIA’s developer blog and its new open models;
• the CUDA toolkit;
• the releases of the engines and tools we run: Ollama, llama.cpp, OpenVINO, PyTorch, ComfyUI, NeMo Relay, PAIR and Claude Code.

A script lists what changed. A research agent reads it and ranks it. R&D checks what matters against the primary source and writes it up here. This is the first edition. It covers the week to 13 October, with 66 new items, of which most were routine.

What we acted on

Claude Code 2.1.291 to 2.1.296: a gap in a sandbox we rely on

Our Atlas experiment (since withdrawn as badly designed) ran Claude Code agents in a locked-down sandbox. A --tools list gives each run exactly four built-in tools. Two fixes in these releases bear directly on that:

• Version 2.1.295 fixed --tools not applying to built-in tools that register after launch.
• Version 2.1.293 stopped Claude being told to use a messaging tool that a --tools list had removed.

The experiment was pinned to an older version, 2.1.197. Our checks read the tool list at start-up, so they could miss a tool that registers later.

What we did:

• The pin moves to 2.1.296 or later before anything runs.
• The harness gains a second check: any tool a run actually uses must be one it was given at the start.

Nothing had run, so nothing was affected. This is exactly the kind of change the watch exists to catch.

Two other changes matter to automated runs:

• one-shot claude -p now waits for background commands instead of cutting them off after five seconds;
• several Bash permission checks were tightened.

NVIDIA PAIR 1.0: context size is now a setting

PAIR spreads model requests across the computers on a home network. We compared it with our own router, interlock, in a brief last week. The 1.0 release notes add three things:

• llama.cpp management, alongside Ollama and LM Studio;
• launch options for each engine, including its context size;
• routing for Claude Code through Ollama and LM Studio, contributed by the community.

The launch options connect with our own measurements. PAIR now lets you set how much context a model gets, but nothing in it tells you how much fits on the card. That is what our E5 experiment measured for six local models on the RTX 5090 (issue 5). We’ve added the release notes to the brief.

NeMo Relay 0.10: traces that stay on the machine

NeMo Relay records what an AI agent actually did, call by call. It’s on our list to trial. Version 0.10 makes that trial simpler:

• It can write traces to local files, so no collector service is needed and nothing leaves the machine.

It also brings one thing to check before a first run:

• It now starts exporting automatically if a standard OpenTelemetry endpoint variable is set in the environment. So the trial starts by confirming no such variable is set.

A new marker shows when Relay let a Claude Code permission request through. That means only that Relay didn’t block it, not that anyone approved it. The release also has breaking changes for existing setups, which don’t affect us, since we haven’t installed it yet.

Worth knowing

• Ollama 0.40.x.

  • Version 0.40.2 upgrades models downloaded by earlier versions the first time they run, and keeps the old copies as backups. That uses extra disk space until they’re removed.
  • Version 0.40.3 paused that background upgrade.
  • Version 0.40.4 is still a pre-release.

  The studio’s 5090 stays on its pinned Ollama while our measurements depend on it. An upgrade will be its own measured step.

• Qwen-Image in NVIDIA’s FP4 format. A 4-bit build of the Qwen-Image model, made for Blackwell GPUs like the RTX 5090.

  • It is published under NVIDIA’s Open Model Agreement. The base model’s Apache 2.0 licence doesn’t carry over.
  • NVIDIA validated it only on data-centre hardware with TensorRT-LLM, and gives no ComfyUI route.

  A candidate to look at, not one we’ve tried.

• NVIDIA’s agent skills catalogue added seven skills, for CodonFM and ontology management. Our mirror of the catalogue is updated.

• Two NVIDIA posts on GPU efficiency:

  • ReplaySSM speeds up decoding for hybrid models at large batch sizes, but not for one user at a time.
  • Green Contexts (CUDA 13.1 and later) divides a GPU between tasks inside one program. It can’t separate two programs, such as Ollama and ComfyUI, which is the case our E4b experiment measured.

Also new this week, and skipped:

• 15 llama.cpp builds;
• 10 NVIDIA posts outside our fields;
• 7 robotics model checkpoints.

From our own bench this week

• The card is full: E5 measured how much context each local model holds on the 5090.
• Multi-repository Atlas: research behind the next step for the studio’s code-map tool, with about thirty sources. An outside model (Kimi K3) reviewed it, writing its own design first. It reached the same design.
• The repository ladder: withdrawn with the Atlas experiment it extended.
• The sources are listed in rnd’s watchlist, each with the reason we follow it.
• rnd watch lists what’s new since the last check, and a research agent reads and ranks it.
• R&D reads the primary source for anything that changes our work before it appears here.

The next watch runs on Monday.