Parameters
2B
Context Length
33K
Modality
Text
Architecture
Dense
KV Cache Rate
48 KB/tok
License
other
Release Date
8 Sept 2026
Knowledge Cutoff
-
API Pricing (per 1M)
Self-hosted only
Precision
Context
Workload Memory: 3.81 GB
INT4 · 8K context
Base Weights: 0.93 GB
KV Cache: 0.38 GB
Runtime & Graphs: 2.50 GB
1x NVIDIA L4 24GB
3.8 / 21.6 GB usable
1x RTX 3060 12GB
3.8 / 10.8 GB usable
Apple M4 24GB
3.8 / 18.0 GB usable
Determine requirements for your specific workload:
Open in VRAM Calculator
Checkpoint format
BF16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 4.40.1
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native BF16 | FlashAttention-3 |
Ada Lovelace, Ampere | Native BF16 | FlashAttention-2 |
No evaluation benchmarks for LFM3 Audio 2B available.
Overall Rank
-
Coding Rank
-
LFM3 Audio 2B is a compact audio-language model developed by Liquid AI for speech perception and sound reasoning. It leverages liquid foundation model dynamics to provide low-latency audio processing on edge hardware.
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
16
Key-Value Heads
4
Attention Head Dimension
128
Position Embedding
ROPE
RoPE Theta
1,000,000
Sliding Window Attention
Yes
Sliding Window Size
256
Sliding Window Ratio
20.8%
Linear Attention
Yes
Linear Attention Ratio
54.2%
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
BF16
Hidden Dimension Size
2,048
Number of Layers
24
FFN Intermediate Size (Dense)
7,680
Tokenizer
Vocabulary Size
128,000
The LFM3 model family developed by Liquid AI.