Skip to content
localmodel.run

Audio model · LFM

LF LFM3 Audio 2B requirements

LFM3 Audio 2B needs about 4.61 GB of memory at Q8_0, audio.cpp ASR, so the lightest hardware that runs it is Nvidia GeForce RTX 2060 (6GB).

Speech recognition · 2B params · Q8_0 GGUF + Q8_0 projector · released Oct 2026.

LFM Open License v1.0 Commercial use: conditional

LFM Open License v1.0: free for organizations below $10M annual revenue; larger organizations require separate terms.

Peak memory (Q8_0, audio.cpp ASR)
~4.61 GB
Runs on CPU
No
Parameters
2B
Type
Speech recognition

Run it

Runtime tools Q8_0, audio.cpp ASR

LFM3 Audio 2B runs in audio.cpp (experimental) at Q8_0, audio.cpp ASR. It needs a GPU (or Apple Silicon with enough unified memory).

audio.cpp (experimental)

Which devices can run LFM3 Audio 2B?

NVIDIA GPUs

Apple Silicon Macs

RAM-only laptops

No mainstream local runtime for LFM3 Audio 2B on RAM-only laptops yet.

iPhone & iPad

No mainstream local runtime for LFM3 Audio 2B on iPhone & iPad yet.

Android

No mainstream local runtime for LFM3 Audio 2B on Android yet.

AMD GPUs

No mainstream local runtime for LFM3 Audio 2B on AMD GPUs yet.

FAQ

How much memory does LFM3 Audio 2B need?

At Q8_0, audio.cpp ASR it consumes ~4.61 GB. It needs a GPU or Apple Silicon.

Can LFM3 Audio 2B run on a phone or CPU?

Not practically. LFM3 Audio 2B expects a GPU; CPU inference is too slow for real use.

Can I use LFM3 Audio 2B commercially?

Conditionally. LFM Open License v1.0: free for organizations below $10M annual revenue; larger organizations require separate terms.

Notes

ASR profile for the experimental audio.cpp runtime with official Q8_0 backbone and Q8_0 projector (2,150,048,544 bytes combined). Runtime-author measurements on 2026-10-27: H100 GPU peak 4,719 MiB for 418-second ASR, used as the memory anchor; short ASR used 2,683 MiB and TTS 2,751 MiB. M3 Ultra process footprint was 3,594 MB for long ASR, a different metric. The anchor applies to that runtime and workload, not all audio lengths or speech tasks. audio.cpp caps context at 8,192 steps, despite the vendor config 128,000-token capacity; Q4 accuracy degradation is documented. CPU and phone memory support are not verified here.

Sources

Memory is a sourced peak-usage anchor at Q8_0, audio.cpp ASR. Catalog updated 2026-11-01. See methodology.