Parameters
260M
Context Length
8K
Modality
Text
Architecture
Dense
KV Cache Rate
23 KB/tok
License
Apache 2.0
Release Date
11 Feb 2025
Knowledge Cutoff
-
API Pricing (per 1M)
Self-hosted only
Precision
Context
Workload Memory: 2.80 GB
INT4 · 8K context
Base Weights: 0.12 GB
KV Cache: 0.18 GB
Runtime & Graphs: 2.50 GB
1x NVIDIA L4 24GB
2.8 / 21.6 GB usable
1x RTX 3060 12GB
2.8 / 10.8 GB usable
Apple M4 24GB
2.8 / 18.0 GB usable
Determine requirements for your specific workload:
Open in VRAM Calculator
Checkpoint format
BF16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 4.47.1
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native BF16 | FlashAttention-3 |
Ada Lovelace, Ampere | Native BF16 | FlashAttention-2 |
No evaluation benchmarks for SmolVLM2 256M Video Instruct available.
Overall Rank
-
Coding Rank
-
SmolVLM2 256M Video Instruct is a lightweight compact multimodal model tailored for video comprehension and instruction following on edge devices. It supports efficient multimodal reasoning with low memory and computational overhead.
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
9
Key-Value Heads
3
Attention Head Dimension
64
Position Embedding
ROPE
RoPE Theta
100,000
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
BF16
Hidden Dimension Size
576
Number of Layers
30
FFN Intermediate Size (Dense)
1,536
Tokenizer
Vocabulary Size
49,280
SmolLM open-weight language models (e.g. SmolLM3)