diff --git a/website/src/components/landing.js b/website/src/components/landing.js
index e84534820..16c342e3f 100644
--- a/website/src/components/landing.js
+++ b/website/src/components/landing.js
@@ -75,14 +75,28 @@ export const LandingBannerGrid = ({ children }) => (
     </Grid>
 )
 
-export const LandingBanner = ({ title, label, to, button, small, background, color, children }) => {
+export const LandingBanner = ({
+    title,
+    label,
+    to,
+    button,
+    small,
+    background,
+    backgroundImage,
+    color,
+    children,
+}) => {
     const contentClassNames = classNames(classes.bannerContent, {
         [classes.bannerContentSmall]: small,
     })
     const textClassNames = classNames(classes.bannerText, {
         [classes.bannerTextSmall]: small,
     })
-    const style = { '--color-theme': background, '--color-back': color }
+    const style = {
+        '--color-theme': background,
+        '--color-back': color,
+        backgroundImage: backgroundImage ? `url(${backgroundImage})` : null,
+    }
     const Heading = small ? H2 : H1
     return (
         <div className={classes.banner} style={style}>
@@ -113,7 +127,7 @@ export const LandingBanner = ({ title, label, to, button, small, background, col
 
 export const LandingBannerButton = ({ to, small, children }) => (
     <div className={classes.bannerButton}>
-        <Button to={to} variant="tertiary" large={!small}>
+        <Button to={to} variant="tertiary" large={!small} className={classes.bannerButtonElement}>
             {children}
         </Button>
     </div>
diff --git a/website/src/images/spacy-irl.jpg b/website/src/images/spacy-irl.jpg
new file mode 100644
index 000000000..ee8f4bdc9
Binary files /dev/null and b/website/src/images/spacy-irl.jpg differ
diff --git a/website/src/styles/landing.module.sass b/website/src/styles/landing.module.sass
index efe3d3e5a..d7340229b 100644
--- a/website/src/styles/landing.module.sass
+++ b/website/src/styles/landing.module.sass
@@ -73,6 +73,7 @@
     color: var(--color-back)
     padding: 5rem
     margin-bottom: var(--spacing-md)
+    background-size: cover
 
 .banner-content
     margin-bottom: 0
@@ -100,7 +101,7 @@
 
 .banner-text-small p
     font-size: 1.35rem
-    margin-bottom: 1rem
+    margin-bottom: 1.5rem
 
 @include breakpoint(min, md)
     .banner-content
@@ -134,6 +135,9 @@
     margin-bottom: var(--spacing-sm)
     text-align: right
 
+.banner-button-element
+    background: var(--color-theme)
+
 .logos
     text-align: center
     padding-bottom: 1rem
diff --git a/website/src/widgets/landing.js b/website/src/widgets/landing.js
index 6905d46d0..9e6e95c2d 100644
--- a/website/src/widgets/landing.js
+++ b/website/src/widgets/landing.js
@@ -19,6 +19,7 @@ import { H2 } from '../components/typography'
 import { Ul, Li } from '../components/list'
 import Button from '../components/button'
 import Link from '../components/link'
+import irlBackground from '../images/spacy-irl.jpg'
 
 import BenchmarksChoi from 'usage/_benchmarks-choi.md'
 
@@ -151,19 +152,21 @@ const Landing = ({ data }) => {
 
             <LandingBannerGrid>
                 <LandingBanner
-                    title="BERT-style language model pretraining and more"
-                    label="New in v2.1"
-                    to="/usage/v2-1"
-                    button="Read more"
+                    title="spaCy IRL 2019: Two days of NLP"
+                    label="Join us in Berlin"
+                    to="https://irl.spacy.io/2019"
+                    button="Get tickets"
+                    background="#ffc194"
+                    backgroundImage={irlBackground}
+                    color="#1a1e23"
                     small
                 >
-                    Learn more from small training corpora by initializing your models with{' '}
-                    <strong>knowledge from raw text</strong>. The new pretrain command teaches
-                    spaCy's CNN model to predict words based on their context, producing
-                    representations of words in contexts. If you've seen Google's BERT system or
-                    fast.ai's ULMFiT, spaCy's pretraining is similar – but much more efficient. It's
-                    still experimental, but users are already reporting good results, so give it a
-                    try!
+                    We're pleased to invite the spaCy community and other folks working on Natural
+                    Language Processing to Berlin this summer for a small and intimate event{' '}
+                    <strong>July 5-6, 2019</strong>. The event includes a hands-on training day for
+                    teams using spaCy in production, followed by a one-track conference. We booked a
+                    beautiful venue, hand-picked an awesome lineup of speakers and scheduled plenty
+                    of social time to get to know each other and exchange ideas.
                 </LandingBanner>
 
                 <LandingBanner
@@ -191,23 +194,17 @@ const Landing = ({ data }) => {
             <LandingLogos title="Featured on" logos={data.logosPublications} />
 
             <LandingBanner
-                title="Convolutional neural network models"
-                label="New in v2.0"
-                button="Download models"
-                to="/models"
+                title="BERT-style language model pretraining"
+                label="New in v2.1"
+                to="/usage/v2-1"
+                button="Read more"
             >
-                spaCy v2.0 features new neural models for <strong>tagging</strong>,{' '}
-                <strong>parsing</strong> and <strong>entity recognition</strong>. The models have
-                been designed and implemented from scratch specifically for spaCy, to give you an
-                unmatched balance of speed, size and accuracy. A novel bloom embedding strategy with
-                subword features is used to support huge vocabularies in tiny tables. Convolutional
-                layers with residual connections, layer normalization and maxout non-linearity are
-                used, giving much better efficiency than the standard BiLSTM solution. Finally, the
-                parser and NER use an imitation learning objective to deliver accuracy in-line with
-                the latest research systems, even when evaluated from raw text. With these
-                innovations, spaCy v2.0's models are <strong>10× smaller</strong>,{' '}
-                <strong>20% more accurate</strong>, and
-                <strong>even cheaper to run</strong> than the previous generation.
+                Learn more from small training corpora by initializing your models with{' '}
+                <strong>knowledge from raw text</strong>. The new pretrain command teaches spaCy's
+                CNN model to predict words based on their context, producing representations of
+                words in contexts. If you've seen Google's BERT system or fast.ai's ULMFiT, spaCy's
+                pretraining is similar – but much more efficient. It's still experimental, but users
+                are already reporting good results, so give it a try!
             </LandingBanner>
 
             <LandingGrid cols={2}>