AI Infra Interviews logo

LFM / Convolution / GQA MoE

LFM2.5-24B-A2B

Active parameters are a compute clue, not a device budget

Open-weight releaseSources checked 2026-11-16

A candidate for private agents when a larger local model is acceptable. Keep its original BF16 checkpoint distinct from the FP8, GGUF and MLX distributions.

24Breported total parameters
2.45Bactive parameters per token
131,072configured token positions

Publisher-reported rounded count; the pinned tensor index is a separate payload measurement.

Evaluate it for

Local reasoning and tool workflows with an artifact-specific serving path.

Choose another path when

Memory favors a much smaller dense checkpoint or the application needs image inputs.

Release context

Why this release matters

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 · Follow the LFM timeline →

Deployment starting points

Which hardware, how many, at what precision?

Start with an exact artifact and an engine that supports it. A configuration below is evidence of a documented path; its device count does not promise a particular throughput or concurrency.

No certified hardware configuration in this edition

Memory favors a much smaller dense checkpoint or the application needs image inputs.

The pinned model card establishes the architecture. It does not, by itself, prove that a particular GPU count runs this checkpoint at your target context. The capacity screen below helps rule out undersized allocations; a successful load and workload test are still required.

47.94 GB of tensor data in the pinned index

This is the index’s reported tensor payload, in decimal GB. It excludes file headers and runtime memory. Mixed precision, conversion and host/device placement determine how much GPU memory the loaded model needs.

Inspect the exact index metadata →
Capacity screen: would the weights fit?

This estimates inference memory for a hypothetical uniform precision. It is useful for rejecting an allocation that is too small. It does not establish a working deployment or estimate training memory.

Weights only, before requests or runtime buffers. These formats are mathematical scenarios, not claims that each artifact exists.
RepresentationApproximate weight floorWhat remains to budget
16-bit uniform48 GBCache, modality state, workspace, quantization metadata and uneven sharding
8-bit uniform24 GBCache, modality state, workspace, quantization metadata and uneven sharding
4-bit uniform12 GBCache, modality state, workspace, quantization metadata and uneven sharding

Enter the total across the whole model replica, not per GPU: KV or recurrent cache for all concurrent requests, image/audio state and execution workspace. There is no universal default. We additionally leave 15% of each GPU unused.

48 GB for hypothetical weights

Enter a request-and-buffer budget to calculate the aggregate capacity requirement.

A count of one does not prove the engine supports this GPU. A count above one does not prove the model shards evenly or communicates efficiently. Publisher-reported rounded count; the pinned tensor index is a separate payload measurement.

Inspect the GPU and its interconnect →

Architecture in practice

What the serving engine has to do

Finite history and full attention have separate budgets.

  1. Short convolution30 layers
    Finite historybounded convolution buffer
  2. Full attention10 layers
    Growing KVretained context history
Short convolution30 layersFinite historybounded convolution bufferFull attention10 layersGrowing KVretained context history

Small active compute does not shrink the expert store.

Count per-request state separately from shared weights.

LFM2.5 schematic, not literal layer order. Thirty short-convolution blocks keep a finite recent history, while ten attention blocks retain full context. Sparse expert weights are an additional shared allocation.

Mechanism schematic based on the pinned model card and configuration. It explains a design principle; it is not a full implementation graph.

Inspect the full checkpoint specifications
Inspected fields for this exact revision. KV heads alone do not describe latent, hybrid or shared-cache layouts.
Exact checkpointLiquidAI/LFM2.5-24B-A2B
Text model typelfm2_moe
Layers40
Hidden width2,048
Attention / KV heads32 / 8
Head dimensionNot verified
Routed / selected experts64 / 4
Configured positions131,072
Documented extensionNot recorded for this checkpoint
Inputs → outputtext → text
License metadataLFM Open License v1.0 · custom terms

30 short convolution layers; 10 full attention layers. These counts describe the configured pattern.

Open weights do not imply unrestricted use. Read the applicable license terms linked from the pinned card.

A useful comparison

What else belongs on the shortlist?

Use the smaller LFM3 for a constrained-memory comparison and Qwen for a larger hybrid-agent comparison; rerun the same tools and stopping rules.

27 Oct 2026

LFM3-1.7B

Small weights can still carry a growing cache

1.73B reported total · text

16 Apr 2026

Qwen3.6-35B-A3B

A smaller expert store for coding-agent experiments

35B reported total · text + image + video

These are editorial comparison candidates. We have not run a matched quality or serving benchmark, so this is not a ranking.

Prepare to explain it

The interview lesson

Reconcile the 24B rounded total, 2.45B active path and 47,943,180,288-byte tensor payload before choosing a device.

Attention layouts · Quantization · Courses and worked examples

For a deployment evaluation, record exact weights, precision, engine, device count, interconnect, prompt/output lengths and concurrency. Report task success, errors, TTFT, TPOT and useful throughput together.

Sources and citation

Our assessment is editorial judgment based on the linked architecture and deployment evidence. Model facts are publisher-reported or attributed to runtime maintainers; no independent GPU benchmark was run.

AI Infra Interviews, “LFM2.5-24B-A2B: release, architecture and deployment”, checked 2026-11-16. Preserve this date and the exact checkpoint when citing.

Model and research notes as JSON · Liquid AI profile