Two vision models make opposite bets: inspectable reasoning or low-latency local work

Moondream 3 Preview and Liquid AI’s LFM2.5-VL-3B take different approaches to image tasks: one can show which parts of an image informed its answer, while the other is designed to respond directly and keep latency low on-device. That distinction is more useful than comparing benchmark scores that do not measure the same thing.

Watch Desk analysis

What happened

An OrcaRouter comparison says Moondream 3 Preview produces image-grounded reasoning traces that refer to specific regions. Its query skill runs in reasoning mode by default, though the article says users can disable that mode. Liquid AI’s LFM2.5-VL-3B is a non-reasoning model, designed for direct answers and lower latency.

The practical hardware gap is striking in OrcaRouter’s account: it says LFM2.5-VL-3B runs in under 3.3GB of memory on an Apple M5 Max, while Moondream 3 Preview has 9 billion parameters in total. Its smaller count of active parameters does not mean the whole model fits into less memory, because the full mixture-of-experts checkpoint must be resident. OrcaRouter says Moondream’s Photon engine supports local use on Apple Silicon Macs and NVIDIA GPUs from the Ampere generation onward.

Why it matters

For a person checking why a vision model flagged a defect or selected an object, an answer that points to relevant image regions could be worth the extra reasoning. For a device that needs fast, direct responses, a smaller model designed for on-device use may be a better fit. Different jobs, different trade-offs. The parameter-count shorthand will not choose for you.

OrcaRouter warns that the headline grounding scores are not a head-to-head comparison: Moondream’s 91.1 RefCOCO result and Liquid AI’s 87.9 RefCOCO precision@1 are different measurements. The figures are vendor-reported, and the article says neither model is available through OrcaRouter’s own service.

Our read

This is a useful comparison because it connects model design to the job and hardware, rather than crowning a winner from incompatible scores. Treat the performance figures as claims to test on your own workload. If you need a human to inspect why the model answered, traceability matters; if you need a quick response on a modest device, latency and memory may matter more.

What to watch

• Whether independent testing compares both models on the same tasks and hardware.
• How much reasoning traces add to response time and token use.
• Whether local deployments can meet the memory and latency needs of real workloads.

Discussion spark: For vision systems whose answers affect a real decision, would you choose a model with inspectable reasoning traces or one built for lower-latency on-device responses?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.