Ollama 0.32.4 fixes Qwen3 MoE decoding and sharpens Apple inference

What changed

ollama/ollama 0.32.4 is a focused local-inference patch with three changes worth separating from the usual release churn.

First, it fixes Qwen3 mixture-of-experts decoding when experts use different quantisation types. It also introduces a faster packed gate/up projection path, with Ollama reporting roughly 4–9% better performance on an M5 Max.

Second, Ollama now quantises the output heads of speculative-decoding draft models at the requested type when creating those drafts. That should reduce unnecessary decode bandwidth while keeping the draft model aligned with the operator’s chosen quantisation.

Third, Laguna models can now run on Apple GPUs through Ollama’s MLX engine. That widens the practical model choices for Apple Silicon machines without requiring a separate inference stack.

Why it matters

The Qwen3 change is primarily a correctness fix. MoE models route each token through selected experts, so mishandling experts quantised in different ways can produce bad decoding long before an operator notices an obvious crash. For agent runtimes such as Hermes and OpenClaw, degraded or inconsistent output is worse than a clean failure: tools may still be selected, but reliability can drift in ways that are difficult to diagnose from infrastructure metrics.

The reported M5 Max gain is useful but not transformative. Four to nine per cent can improve latency and throughput on a machine already serving local models, but it does not rescue a model that is too large, too slow or too weak at tool use. Treat it as a welcome implementation gain rather than a reason to change hardware.

The draft-model change matters to anyone experimenting with speculative decoding. Drafts only help when their extra work is cheaper than waiting for the main model. Quantising output heads at the requested type attacks part of that bandwidth cost, but the end-to-end benefit still depends on acceptance rate, model pairing and prompt shape.

MLX support for Laguna is operationally useful for Foundry’s Apple-based local evaluation. It makes another model family available inside the same Ollama service boundary used for other tests, reducing the integration work needed before comparing model quality.

My read

Worth a targeted update, not an estate-wide rush. If an Ollama node serves Qwen3 MoE, uses speculative-decoding drafts or needs Laguna on Apple Silicon, move to 0.32.4 after a short smoke test. The mixed-quantisation fix alone justifies the patch for affected Qwen3 deployments.

Validate with the actual quantised model and workload:

1. replay a small set of known tool-calling conversations;
2. compare output quality and tool selection against 0.32.3;
3. record first-token latency and tokens per second;
4. for speculative decoding, measure acceptance rate as well as raw throughput;
5. on Apple Silicon, confirm the intended model is using MLX and is not silently falling back to another path.

If none of those paths are in use, this can wait for the normal maintenance window. There is no broad agent-runtime security or API compatibility issue in the published notes.

Bottom line

Ollama 0.32.4 is small but material for a specific local-inference slice: Qwen3 MoE decoding becomes safer, speculative drafts get a more appropriate quantisation path, and Laguna reaches Apple GPUs through MLX. Update affected systems and benchmark the result. Everyone else can watch rather than manufacture urgency.