AI Infra Interviews logo

LFM / Lab profile

Liquid AI

Local agents still need a complete state budget.

These checkpoints provide a smaller-device comparison to the frontier cluster releases. LFM2.5-24B-A2B uses sparse experts; LFM3-1.7B uses a much smaller dense model with three kinds of sequence-processing layers.

Architecture trajectory

What connects these models?

Short convolutions retain finite history. Sliding attention retains a bounded window when the implementation reclaims it. Full attention keeps growing history. Sum the appropriate state for each family.

Three layer families keep different history.

  1. Short convolution13 layers
    Finite historybounded convolution buffer
  2. Sliding attention5 layers; window 256
    Local KVreclaim outside the window
  3. Full attention6 layers
    Growing KVretained context history
Short convolution13 layersFinite historybounded convolution bufferSliding attention5 layers; window 256Local KVreclaim outside the windowFull attention6 layersGrowing KVretained context history

Convolution history is not an attention KV array.

Global layers keep growing after local windows fill.

LFM3 schematic, not literal layer order. Thirteen short convolutions retain finite history; five attention layers use a local window; six attention layers retain full history. Reclaiming local state does not bound the full-attention cache.

Newest first / grouped sequence, not a time scale

The LFM timeline

  1. LFM3-1.7B

    Small weights can still carry a growing cache

    The 24-layer stack has 13 short convolutions, five sliding-attention layers and six full-attention layers. Only the local portions can stop growing with context.

    Publisher release post; repository creation predates the announcement.

    Release evidence ↗

  2. LFM2.5-24B-A2B

    Active parameters are a compute clue, not a device budget

    The 40-layer design combines thirty short convolutions with ten grouped-query attention layers. Routing selects four of 64 experts; their stored tensors remain part of residency.

    Publisher release post; not the earlier repository creation date.

    Release evidence ↗

Which other labs should you compare?

Compare with small Qwen or Gemma checkpoints under the same local task and memory limit, including output length and tool parsing.

Qwen

Alibaba Qwen

One family spans tiny dense models, coding MoEs and conditional memory.

Gemma

Google DeepMind

Gemma is a family of different design choices, not just different sizes.

Interview preparation

What you should be able to explain

Explain why a 1.73B model can still use 384 MiB for its full-attention KV at 32,768 positions, before other state.

Study the underlying systems → · Get the illustrated report →

Sources and scope

Covered checkpoints carry their own research dates. Assessments and comparison candidates are our interpretation of the cited designs, not measured cross-lab rankings. Each timeline entry links to its release evidence; model pages preserve pinned card and configuration sources.