AI Infra Interviews logo

LFM / Convolution / sliding / full attention

LFM3-1.7B

Small weights can still carry a growing cache

Open-weight releaseSources checked 2026-11-16

Consider this checkpoint for constrained text workflows and local tool use. It is a deployment contrast to the large cluster models, not a claim of equivalent frontier reasoning quality.

1.73Breported total parameters
Not verifiedactive parameters per token
262,144configured token positions

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

Evaluate it for

Task-specific local agents after evaluation on the intended tools and language.

Choose another path when

Complex coding or broad knowledge is the main requirement, or the license does not permit the intended use.

Release context

Why this release matters

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 · 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

Complex coding or broad knowledge is the main requirement, or the license does not permit the intended use.

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.

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 uniform3.46 GBCache, modality state, workspace, quantization metadata and uneven sharding
8-bit uniform1.73 GBCache, modality state, workspace, quantization metadata and uneven sharding
4-bit uniform0.865 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.

3.46 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

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.

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/LFM3-1.7B
Text model typelfm3
Layers24
Hidden width2,048
Attention / KV heads16 / 4
Head dimension128
Routed / selected expertsNot verified / Not verified
Configured positions262,144
Documented extensionNot recorded for this checkpoint
Inputs → outputtext → text
License metadataLFM Open License v1.0 · custom terms

13 short convolution layers; 5 sliding attention layers; 6 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?

The 24B sibling trades a larger expert store for a different capability and state profile. Compare task success and peak memory on the actual device.

13 Oct 2026

LFM2.5-24B-A2B

Active parameters are a compute clue, not a device budget

24B reported total · text

2 Mar 2026

Qwen3.5-4B

A practical small-model starting point

4B 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

At two-byte KV, derive 2 KiB per attention-layer position. At 32,768 positions the six full layers alone retain 384 MiB before convolution state and runtime overhead.

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, “LFM3-1.7B: 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