Hy-Memory and Liquid AI’s model show why sparse does not mean small

Two sparse AI models make very different bets on where they should run: Tencent’s Hy-Memory is presented as a memory-writing component for a server, while Liquid AI’s LFM2.5 is built for on-device agents. The practical divide is in the files: the comparison lists Hy-Memory at 60.1 GB, against a 14.5 GB quantised download for Liquid’s model.

Watch Desk analysis

What happened

An OrcaRouter analysis compares the models’ architecture, packaging and intended jobs. It says Hy-Memory-30B-A3B has 30 billion parameters, about three billion active per token, and is released as a 60.1 GB BF16 checkpoint. The article says no quantised alternative was available in the release it examined.

Liquid AI’s LFM2.5-24B-A2B is listed at 24 billion parameters, with 2.45 billion active per token. OrcaRouter lists a 14.5 GB Q4KM GGUF and a 13.6 GB quantisation-aware version, alongside the full-precision weights. It describes Liquid’s model as an agent for local use, and Hy-Memory as a component that turns conversations into structured memory records.

Key findings

• Sparse does not mean tiny
  Active parameter counts describe what is used per token, not the size of the full checkpoint that must be loaded.
• The available formats change the hardware question
  The comparison lists smaller quantised files for Liquid, while Hy-Memory’s listed BF16 checkpoint is 60.1 GB.
• The models are aimed at different jobs
  Liquid is presented as an agent that plans and calls tools; Hy-Memory writes structured memories for later use.

Why it matters

A model name’s active-parameter count can make local deployment sound easier than it is. The full checkpoint still matters for memory, and available quantisations can change whether a model is practical on hardware someone already owns. That makes the release formats part of the product story, not just packaging trivia.

The comparison also cautions against treating these models as straightforward rivals. One is described as an agent, the other as a memory writer. Teams building an AI system might consider them for different parts of a pipeline, rather than choosing one winner from a leaderboard.

Our read

The useful takeaway is less “which model wins?” and more “what does the job require, and can your hardware load the thing?” OrcaRouter’s figures and descriptions make that comparison concrete, but they are the analysis’s account, not independent benchmark results. Check the model cards, licence and actual hardware needs before planning a deployment; sparse arithmetic is no substitute for free memory.

What to watch

• Whether Hy-Memory gains quantised versions that reduce its download and memory footprint.
• Whether independent users reproduce the reported model characteristics and test the intended workloads.
• Whether developers use a dedicated memory-writing model alongside an agent, or prefer a single model to do both.

Discussion spark: Would you rather run separate models for memory and tool use, or keep an AI system simpler with one model doing both?

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.