Discussion

Liquid AI’s new model and Gemma 4 make similar compute trade-offs, but very different ones

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#5659

Liquid AI’s new LFM2.5-24B-A2B and Google DeepMind’s Gemma 4 26B-A4B have similar headline parameter counts, but very different strengths, memory demands and use cases. A detailed comparison from OrcaRouter offers practical guidance for choosing between the open-weight models, with an important caveat: its benchmark results come from Liquid AI, not independent testing.

Watch Desk analysis

What happened

The comparison puts Liquid’s 24-billion-parameter model, with 2.45 billion active per token, alongside Gemma 4’s 25.2 billion parameters and 3.8 billion active per token. Both are mixtures of experts, but the resemblance stops there. OrcaRouter says Liquid’s model is text-only, geared towards reasoning and tool use, and offers a 131,072-token context. Gemma 4 supports images and a 262,144-token context.

The practical distinction is storage. OrcaRouter lists Liquid’s Q4KM model file at 14.5 GB and its BF16 version at 48 GB. It gives Gemma 4’s BF16 weights as roughly 50 GB, before accounting for its separate multimodal projection file. The comparison also notes that sparse models still need their full checkpoint resident in memory: fewer active parameters can help throughput, but do not make a large model small enough for a modest-memory machine.

Key findings

  • Liquid is the more compact download
    OrcaRouter lists its Q4KM file at 14.5 GB, compared with about 50 GB for Gemma 4’s BF16 weights.
  • Gemma sees images and handles longer context
    Gemma 4 supports image input and a 262,144-token context; Liquid’s model is text-only with a 131,072-token context.
  • The published comparison favours Liquid on several agent tasks
    Liquid’s table reports higher scores than Gemma 4 on benchmarks including AutomationBench and ToolSandbox, while Gemma leads on PinchBench and SciCode.
  • The benchmark magnitudes deserve caution
    Liquid ran the models in a shared vLLM setup, but the results and table are vendor-reported, and the comparison flags differences in prompts and generation settings.

Why it matters

“Active parameters” are easy to mistake for the memory needed to run a model. In this comparison, both checkpoints are around 25 billion parameters, even though only a smaller fraction is active per token. For anyone trying to run models locally, the file sizes and supported hardware matter more than the neat arithmetic in the model names.

The comparison also shows why “which model is better?” is the wrong first question. Liquid’s model is aimed at text, reasoning and tool use; Gemma offers image input, a longer context and an Apache 2.0 licence. A model that cannot process images is not a contender for an image-heavy workflow, however appealing its agent benchmark scores may look.

Our read

This is a useful buying and deployment comparison, not a neutral leaderboard. OrcaRouter gives readers concrete differences to weigh, while Liquid’s own benchmark table is evidence of what the company tested and reported, not a settled ranking. Start with the task, then check the checkpoint size against the memory you actually have. The numbers in a model’s name are not a RAM budget, however much we might wish they were.

What to watch

  • Whether independent tests reproduce the reported agent-task results.
  • How the models perform on readers’ own tasks and hardware.
  • Whether quantised formats change the memory and performance trade-offs.
  • Whether later versions add capabilities, context or licensing changes that alter the choice.

Discussion spark: For a local agent, would you choose Liquid’s smaller text-focused model for tool use, or Gemma 4’s image support and longer context, even with its larger memory footprint?

Sources and evidence

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.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.