diff --git a/candle-examples/examples/beit/README.md b/candle-examples/examples/beit/README.md
new file mode 100644
index 0000000000..23af1e32f5
--- /dev/null
+++ b/candle-examples/examples/beit/README.md
@@ -0,0 +1,20 @@
+# candle-beit
+
+[Beit](https://arxiv.org/abs/2106.08254) is a computer vision model.
+In this example, it is used as an ImageNet classifier: the model returns the
+probability for the image to belong to each of the 1000 ImageNet categories.
+
+## Running some example
+
+```bash
+cargo run --example beit --release -- --image candle-examples/examples/yolo-v8/assets/bike.jpg
+
+> mountain bike, all-terrain bike, off-roader: 56.16%
+> bicycle-built-for-two, tandem bicycle, tandem: 3.08%
+> maillot                 : 2.23%
+> alp                     : 0.88%
+> crash helmet            : 0.85%
+
+```
+
+![Leading group, Giro d'Italia 2021](../yolo-v8/assets/bike.jpg)
diff --git a/candle-examples/examples/beit/main.rs b/candle-examples/examples/beit/main.rs
new file mode 100644
index 0000000000..5ef2a6ae0c
--- /dev/null
+++ b/candle-examples/examples/beit/main.rs
@@ -0,0 +1,79 @@
+//! BEiT: BERT Pre-Training of Image Transformers
+//! https://github.com/microsoft/unilm/tree/master/beit
+
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use clap::Parser;
+
+use candle::{DType, Device, IndexOp, Result, Tensor, D};
+use candle_nn::{Module, VarBuilder};
+use candle_transformers::models::beit;
+
+/// Loads an image from disk using the image crate, this returns a tensor with shape
+/// (3, 384, 384). Beit special normalization is applied.
+pub fn load_image384_beit_norm<P: AsRef<std::path::Path>>(p: P) -> Result<Tensor> {
+    let img = image::io::Reader::open(p)?
+        .decode()
+        .map_err(candle::Error::wrap)?
+        .resize_to_fill(384, 384, image::imageops::FilterType::Triangle);
+    let img = img.to_rgb8();
+    let data = img.into_raw();
+    let data = Tensor::from_vec(data, (384, 384, 3), &Device::Cpu)?.permute((2, 0, 1))?;
+    let mean = Tensor::new(&[0.5f32, 0.5, 0.5], &Device::Cpu)?.reshape((3, 1, 1))?;
+    let std = Tensor::new(&[0.5f32, 0.5, 0.5], &Device::Cpu)?.reshape((3, 1, 1))?;
+    (data.to_dtype(candle::DType::F32)? / 255.)?
+        .broadcast_sub(&mean)?
+        .broadcast_div(&std)
+}
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    image: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    let device = candle_examples::device(args.cpu)?;
+
+    let image = load_image384_beit_norm(args.image)?.to_device(&device)?;
+    println!("loaded image {image:?}");
+
+    let model_file = match args.model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model("vincent-espitalier/candle-beit".into());
+            api.get("beit_base_patch16_384.in22k_ft_in22k_in1k_adapted.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+    let model = beit::vit_base(vb)?;
+    println!("model built");
+    let logits = model.forward(&image.unsqueeze(0)?)?;
+    let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
+        .i(0)?
+        .to_vec1::<f32>()?;
+    let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
+    prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
+    for &(category_idx, pr) in prs.iter().take(5) {
+        println!(
+            "{:24}: {:.2}%",
+            candle_examples::imagenet::CLASSES[category_idx],
+            100. * pr
+        );
+    }
+    Ok(())
+}
diff --git a/candle-transformers/src/models/beit.rs b/candle-transformers/src/models/beit.rs
new file mode 100644
index 0000000000..c534032cc2
--- /dev/null
+++ b/candle-transformers/src/models/beit.rs
@@ -0,0 +1,367 @@
+use candle::{DType, IndexOp, Result, Tensor, D};
+use candle_nn::{layer_norm, LayerNorm, Linear, Module, VarBuilder};
+
+const IMG_SIZE: usize = 384;
+const PATCH_SIZE: usize = 16;
+const NUM_CLASSES: usize = 1000;
+const WINDOW_SIZE: usize = IMG_SIZE / PATCH_SIZE; // 384 / 16 = 24
+const NB_TOKENS: usize = WINDOW_SIZE * WINDOW_SIZE + 1; // 24 * 24 + 1 = 577
+
+fn linear(vb: VarBuilder, in_dim: usize, out_dim: usize, bias: bool) -> Result<Linear> {
+    if bias {
+        candle_nn::linear(in_dim, out_dim, vb)
+    } else {
+        candle_nn::linear_no_bias(in_dim, out_dim, vb)
+    }
+}
+
+#[derive(Debug)]
+struct Attention {
+    qkv: Linear,
+    proj: Linear,
+    relative_position_bias_table: Tensor,
+    relative_position_index: Tensor,
+    num_heads: usize,
+    scale: f64,
+}
+
+impl Attention {
+    fn new(
+        vb: VarBuilder,
+        dim: usize,
+        num_heads: usize,
+        qkv_bias: bool,
+        proj_bias: bool,
+        relative_position_index: &Tensor,
+    ) -> Result<Self> {
+        let qkv = linear(vb.pp("qkv"), dim, dim * 3, qkv_bias)?;
+        let proj = linear(vb.pp("proj"), dim, dim, proj_bias)?;
+        // num_relative_distance = token-token(47x47) + token-CLS(1) + CLS-token(1) + CLS-CLS(1) = 2212
+        let num_relative_distance = (2 * WINDOW_SIZE - 1) * (2 * WINDOW_SIZE - 1) + 3;
+        let relative_position_bias_table = vb.get(
+            (num_relative_distance, num_heads),
+            "relative_position_bias_table",
+        )?;
+        let relative_position_index = relative_position_index.clone();
+        let scale = 1. / ((dim / num_heads) as f64).sqrt();
+        Ok(Self {
+            qkv,
+            proj,
+            relative_position_bias_table,
+            relative_position_index,
+            num_heads,
+            scale,
+        })
+    }
+}
+
+impl Attention {
+    fn _get_rel_pos_bias(&self) -> Result<Tensor> {
+        self.relative_position_bias_table
+            .index_select(
+                &self
+                    .relative_position_index
+                    .flatten_all()?
+                    .to_dtype(DType::U32)?,
+                0,
+            )?
+            .reshape((NB_TOKENS, NB_TOKENS, ()))?
+            .transpose(0, 1)? // 102
+            .transpose(0, 2)? // 201
+            .contiguous()?
+            .unsqueeze(0)
+    }
+}
+
+impl Module for Attention {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let (b, n, c) = xs.dims3()?;
+        let qkv = self
+            .qkv
+            .forward(xs)?
+            .reshape((b, n, 3, self.num_heads, c / self.num_heads))?
+            .transpose(1, 2)? // 02134
+            .transpose(0, 1)? // 20134
+            .transpose(2, 3)?; // 20314
+        let q = (qkv.i(0)? * self.scale)?;
+        let k = qkv.i(1)?.contiguous()?;
+        let v = qkv.i(2)?.contiguous()?;
+        let attn = (&q.matmul(&k.t()?)? + self._get_rel_pos_bias())?;
+        let attn = candle_nn::ops::softmax(&attn, D::Minus1)?;
+        let attn = attn.matmul(&v)?.transpose(1, 2)?.reshape((b, n, c))?;
+        self.proj.forward(&attn)
+    }
+}
+
+#[derive(Debug)]
+struct LayerScale {
+    gamma: Tensor,
+}
+
+impl LayerScale {
+    fn new(vb: VarBuilder, dim: usize) -> Result<Self> {
+        let gamma = vb.get(dim, "gamma")?;
+        Ok(Self { gamma })
+    }
+}
+
+impl Module for LayerScale {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        xs.broadcast_mul(&self.gamma)
+    }
+}
+
+#[derive(Debug)]
+struct Mlp {
+    fc1: Linear,
+    fc2: Linear,
+}
+
+impl Mlp {
+    fn new(vb: VarBuilder, in_features: usize, hidden_features: usize, bias: bool) -> Result<Self> {
+        let out_features = in_features;
+        let fc1 = linear(vb.pp("fc1"), in_features, hidden_features, bias)?;
+        let fc2 = linear(vb.pp("fc2"), hidden_features, out_features, bias)?;
+        Ok(Self { fc1, fc2 })
+    }
+}
+
+impl Module for Mlp {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let xs = self.fc1.forward(xs)?.gelu()?;
+        self.fc2.forward(&xs)
+    }
+}
+
+#[derive(Debug)]
+struct Block {
+    norm1: LayerNorm,
+    attn: Attention,
+    ls1: LayerScale,
+    norm2: LayerNorm,
+    mlp: Mlp,
+    ls2: LayerScale,
+}
+
+impl Block {
+    fn new(
+        vb: VarBuilder,
+        dim: usize,
+        num_heads: usize,
+        relative_position_index: &Tensor,
+    ) -> Result<Self> {
+        let norm1 = layer_norm(dim, 1e-6, vb.pp("norm1"))?;
+        let attn = Attention::new(
+            vb.pp("attn"),
+            dim,
+            num_heads,
+            true,
+            true,
+            relative_position_index,
+        )?;
+        let ls1 = LayerScale::new(vb.pp("ls1"), dim)?;
+        let norm2 = layer_norm(dim, 1e-6, vb.pp("norm2"))?;
+        let mlp = Mlp::new(vb.pp("mlp"), dim, dim * 4, true)?;
+        let ls2 = LayerScale::new(vb.pp("ls2"), dim)?;
+        Ok(Self {
+            norm1,
+            attn,
+            ls1,
+            norm2,
+            mlp,
+            ls2,
+        })
+    }
+}
+
+impl Module for Block {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let residual = xs;
+        let xs = self
+            .ls1
+            .forward(&self.attn.forward(&self.norm1.forward(xs)?)?)?;
+        let xs = (xs + residual)?;
+        let residual = &xs;
+        let xs = self
+            .ls2
+            .forward(&self.mlp.forward(&self.norm2.forward(&xs)?)?)?;
+        xs + residual
+    }
+}
+
+#[derive(Debug)]
+struct PatchEmbed {
+    proj: candle_nn::Conv2d,
+    patch_size: (usize, usize),
+}
+
+impl PatchEmbed {
+    fn new(vb: VarBuilder, patch_size: usize, in_chans: usize, embed_dim: usize) -> Result<Self> {
+        let config = candle_nn::Conv2dConfig {
+            stride: patch_size,
+            ..Default::default()
+        };
+        let proj = candle_nn::conv2d(in_chans, embed_dim, patch_size, config, vb.pp("proj"))?;
+        Ok(Self {
+            proj,
+            patch_size: (patch_size, patch_size),
+        })
+    }
+}
+
+impl Module for PatchEmbed {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let (_b, _c, h, w) = xs.dims4()?;
+        let (patch_h, patch_w) = self.patch_size;
+        if (h % patch_h) != 0 {
+            candle::bail!("image height {h} is not a multiple of patch height {patch_h}")
+        }
+        if (w % patch_w) != 0 {
+            candle::bail!("image width {w} is not a multiple of patch width {patch_w}")
+        }
+        let xs = self.proj.forward(xs)?;
+        let (b, c, h, w) = xs.dims4()?;
+        // flatten embeddings.
+        xs.reshape((b, c, h * w))?.transpose(1, 2)
+    }
+}
+
+#[derive(Debug)]
+pub struct BeitVisionTransformer {
+    patch_embed: PatchEmbed,
+    cls_token: Tensor,
+    blocks: Vec<Block>,
+    norm: LayerNorm,
+    head: Linear,
+}
+
+impl BeitVisionTransformer {
+    pub fn new(vb: VarBuilder, depth: usize, embed_dim: usize, num_heads: usize) -> Result<Self> {
+        let patch_embed = PatchEmbed::new(vb.pp("patch_embed"), PATCH_SIZE, 3, embed_dim)?;
+        let cls_token = vb.get((1, 1, embed_dim), "cls_token")?;
+        let head = linear(vb.pp("head"), embed_dim, NUM_CLASSES, true)?;
+        let relative_position_index = vb.get((NB_TOKENS, NB_TOKENS), "relative_position_index")?;
+        let norm = layer_norm(embed_dim, 1e-6, vb.pp("norm"))?;
+        let vb_b = vb.pp("blocks");
+        let blocks = (0..depth)
+            .map(|i| {
+                Block::new(
+                    vb_b.pp(&i.to_string()),
+                    embed_dim,
+                    num_heads,
+                    &relative_position_index,
+                )
+            })
+            .collect::<Result<Vec<_>>>()?;
+        Ok(Self {
+            patch_embed,
+            cls_token,
+            blocks,
+            norm,
+            head,
+        })
+    }
+
+    fn prepare_tokens_with_mask(&self, xs: &Tensor) -> Result<Tensor> {
+        let xs = self.patch_embed.forward(xs)?;
+        Tensor::cat(&[&self.cls_token, &xs], 1)
+    }
+
+    fn get_intermediate_layers_not_chunked(
+        &self,
+        xs: &Tensor,
+        blocks_to_take: &[usize],
+    ) -> Result<Vec<Tensor>> {
+        let mut xs = self.prepare_tokens_with_mask(xs)?;
+        let mut output = Vec::new();
+        for (i, blk) in self.blocks.iter().enumerate() {
+            xs = blk.forward(&xs)?;
+            if blocks_to_take.contains(&i) {
+                output.push(xs.clone());
+            }
+        }
+        if output.len() != blocks_to_take.len() {
+            candle::bail!(
+                "only {} / {} blocks found",
+                output.len(),
+                blocks_to_take.len()
+            );
+        }
+        Ok(output)
+    }
+
+    pub fn get_intermediate_layers(
+        &self,
+        xs: &Tensor,
+        blocks_to_take: &[usize],
+        reshape: bool,
+        return_class_token: bool,
+        norm: bool,
+    ) -> Result<Tensor> {
+        let outputs = self.get_intermediate_layers_not_chunked(xs, blocks_to_take)?;
+        let outputs = if norm {
+            outputs
+                .iter()
+                .map(|out| self.norm.forward(out))
+                .collect::<Result<Vec<_>>>()?
+        } else {
+            outputs
+        };
+        let class_tokens = outputs
+            .iter()
+            .map(|out| out.i((.., 0)))
+            .collect::<Result<Vec<_>>>()?;
+        let outputs = outputs
+            .iter()
+            .map(|out| out.i((.., 1..)))
+            .collect::<Result<Vec<_>>>()?;
+
+        let outputs = if reshape {
+            let (b, _c, w, h) = xs.dims4()?;
+            let patch_size = self.patch_embed.patch_size.0;
+            let num_channels = outputs[0].elem_count() / (b * (w / patch_size) * (h / patch_size));
+            outputs
+                .iter()
+                .map(|out| {
+                    out.reshape((b, w / patch_size, h / patch_size, num_channels))?
+                        .transpose(2, 3)?
+                        .transpose(1, 2)
+                })
+                .collect::<Result<Vec<_>>>()?
+        } else {
+            outputs
+        };
+
+        let outputs = if return_class_token {
+            outputs
+                .iter()
+                .zip(class_tokens.iter())
+                .map(|(out, class_token)| Tensor::cat(&[out, class_token], D::Minus1))
+                .collect::<Result<Vec<_>>>()?
+        } else {
+            outputs
+        };
+
+        Tensor::stack(&outputs[..], 0)
+    }
+}
+
+impl Module for BeitVisionTransformer {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let mut xs = self.prepare_tokens_with_mask(xs)?;
+        for blk in self.blocks.iter() {
+            xs = blk.forward(&xs)?
+        }
+        let xs_moy_local_tokens = xs.i((.., 1..))?.mean(1)?;
+        let xs_norm = self.norm.forward(&xs_moy_local_tokens)?;
+        self.head.forward(&xs_norm)
+    }
+}
+
+pub fn vit_base(vb: VarBuilder) -> Result<BeitVisionTransformer> {
+    BeitVisionTransformer::new(vb, 12, 768, 12)
+}
+
+pub fn vit_large(vb: VarBuilder) -> Result<BeitVisionTransformer> {
+    BeitVisionTransformer::new(vb, 24, 1024, 16)
+}
diff --git a/candle-transformers/src/models/mod.rs b/candle-transformers/src/models/mod.rs
index 2908d3457a..d95d30ae5a 100644
--- a/candle-transformers/src/models/mod.rs
+++ b/candle-transformers/src/models/mod.rs
@@ -1,3 +1,4 @@
+pub mod beit;
 pub mod bert;
 pub mod bigcode;
 pub mod blip;