Ollama 0.32.3 fixes dropped tool calls and agent integration regressions

What changed

ollama/ollama 0.32.3 is a patch release covering the changes since 0.32.1. The important part is not a new model badge. It is a cluster of fixes around the plumbing used by coding agents and tool-calling workflows:

• Claude Code Channels have been restored after an integration regression.
• Anthropic thinking streams have been fixed.
• GLM tool calls that could be silently dropped at the end of generation are now preserved.
• Hermes Desktop now respects --force-build.
• model downloads that stalled before sending data have been fixed.
• MLX and llama.cpp engines have been updated.

The release also broadens GPU support with CUDA on Windows ARM64, B200 support through CUDA 12, and lower memory use on Linux CUDA/ROCm integrated GPUs. Laguna 2.1 models gain chat, thinking and tool-calling support, alongside a Metal inference fix.

A separate default-branch change after the release changes model creation so lm_head weights use the requested quantisation family’s aligned 8-bit type. That aims to retain quality near BF16 while reducing decode bandwidth, but it is not part of the 0.32.3 tag.

Why it matters

Tool-call transport is operational infrastructure. If a normal text response is malformed, a user notices. If the final tool call disappears, an agent can finish with a confident answer while the underlying action never happened. That is the nastier failure mode because dashboards can remain green and the conversation can still look plausible.

The same applies to thinking streams and channel integrations. Hermes and OpenClaw may not use every Ollama integration directly, but they sit in the same class of systems: multi-step runtimes that need model output, stream events and requested actions to survive the trip between model and executor without being reordered, truncated or dropped.

The Hermes Desktop fix is particularly direct. Respecting --force-build makes explicit operator intent reliable again, which matters during local development and when diagnosing stale builds.

The hardware changes are useful but secondary for most existing Foundry deployments. They widen where Ollama can run efficiently; they do not, by themselves, justify moving a stable service to unfamiliar hardware.

My read

Update now — with a short agent smoke test. This is not an emergency security patch, but it fixes exactly the sort of silent integration failures that are expensive to diagnose later.

For an Ollama-backed agent service, the sensible validation is small and concrete:

1. run a prompt that emits one tool call at the end of generation;
2. confirm the executor receives and records it exactly once;
3. test a streamed reasoning response without malformed or missing events;
4. download a small uncached model to verify transfer starts normally;
5. if using Hermes Desktop, confirm --force-build actually rebuilds the intended artefacts.

Pin 0.32.3 rather than following the default branch. The later lm_head quantisation change is interesting for locally created low-bit models, but it should be benchmarked for output quality and decode speed before adoption.

Bottom line

Ollama 0.32.3 is a boring release in the best sense: fewer stalled downloads, fewer broken integrations and, most importantly, fewer agent actions vanishing at the end of generation. Update Ollama-backed agent environments after a focused smoke test; watch the newer 8-bit lm_head work separately rather than conflating it with the release.