Slides: The Small Model Infrastructure Nobody Built (So We Did) — Filip Makraduli, Superlinked
Source Video
The Small Model Infrastructure Nobody Built (So We Did) — Filip Makraduli, Superlinked
Relationship To World's Fair 2026
These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material.
Related Scheduled Sessions
- No individual scheduled session mapping has been assigned yet; treat this as an event livestream deck.
Extracted Slides

OCR text:
PLATINUM SPONSORS
Braintrust WorkOS OpenAI

OCR text:
TheSmall-Model
Infrastructure
Nobody Built
(sowe did)
FilipMakraduli·Superlinked
AlEngineer
EUROPE

OCR text:
A WRITER'S DIARY ON AI
What Actually Makes Embedding
Model Inference Fast?
From Flash Attention to Quantization, where is the inference bottleneck? Is it the architecture, the maths, or will writing everything in Rust solve all my problems? (Hint: It's Not Rust)
FILIP MAKRADULI
JAN 22, 2026
Not quite.
Google DeepMind

OCR text:
I realised... I had to learn
AI Engineer
EUROPE

OCR text:
Traming & re tening area pose DCMT RL stenet a gern
va Eva‘vation & bencnmark:ng GPU uti! zat on & schedu .ng
* ~ Appied Al researc Rosting & autosca’ ng
# . Bem rae feat ien Crployment & oprrat ons
bod *
ad os *
Best way to learn?
Join the team building it.
: | Al Engineer |
eS) EUROPE
=

OCR text:
.
Superlinked ©) github.com/superlinked/sie
rane
Ps * VC - backed Al infrastructure company building the model stack for Al search and document
re x processing
ve
Works with your favorite taols neers
ey Pa Pree’ os Pera te} coy paren ae
Cree nears oes ee ere vena Sone Un Tn eee earns
eer et Beene eee es prereset an
Cate iat tne ie Ld Cha Se Ct we aac d oe eon a ee Lael pee OE i rec
teat eer eaed ene Learns] coe ae ar ar Ree eee Tre el
eee te Cm Me eerie seer ie at et ere de he Pe A ae ee
Fo ett loses Pe oes re
l | Al Engineer |
= aUelaa
re = oF

OCR text:
PU emulate uh
Performance degrades as contest fille up
lel Ret CECE LOE CDC
aa _ / vere Tee ah ee Mae Lis
Oa a TM, . 2a,
* a N atl eeelre)
rs N
4 : N\
: \- . Es
: . “iy . "Why not just Opus 4.62"
. = Ty, se!
: | Al Engineer |
eS See) aa
a hati

OCR text:
Context management
Taxonomy classification
1OK categories Shopify product catalogue text + images
Pan 3
Pu *
aCe
Sg a Tool calling
ars Product
bauer sO) a ae
Switching model type = changing one parameter, not
rebuilding infra
] Engineering the future of Al

OCR text:
Model Performance V5 Size
we _ ‘ :
te ey °
Y
one ”
Pane d
ne nom &
a
* ad ‘ e °
a in .
2
.
Pole Muitel See titan Paraet ona)
| Al Engineer |
Eng EUROPE
aoe

OCR text:
SO, LET'S SUPPORT 85+ MODELS WITH 35+ RUNTIMES AND INFERENCE BACKENDS
Every architecture is different under the hood
bert_flash
Fixed position IDs (0,1,2,...)
Each token knows its absolute slot
qwen2_flash
RoPE: rotates Q/K vectors by position
GQA: fewer KV heads = less memory
modernbert_flash
Fused QKV: one matrix, not three
Pre-normalize before attention
colbert_*
Keeps every token's vector (not pooled)
Late interaction: match token-by-token
cross_encoder
Takes a pair (query + doc) as input
Outputs a score, not a vector
clip_siglip_florence2
Image encoder + text encoder
Shared embedding space
splade_flash
Sparse 250K-dim vocabulary vector
Learned term weights, not dense floats
sglang
48-88 embedding models
KV cache = paged memory
No universal engine
BERT + Qwen2 + ModernBERT.
Each needs its own inference path.
An agentic workflow
We wrote each adapter using AI agents
until we had coverage across 90 models.
Re-implementation
Not wrappers. We re-implement
the forward pass and pooling for each arch.
15 of 35 adopters use Flash Attention 2 with variable-length sequence packing

OCR text:
cee Qwen2 ore a1 as
Norm Pre-norm RMSNorm
é Ca i‘ Q/KIV Rate Separate (GQA)
na ae
* rar i Lexth ated) Coa
Attention GGA (fewer KV)
Output
Allon packed sequences with = zero padding waste
Engineering the future of Al
ES) >

OCR text:
We open sourced both the yin (model inference) and the
4, a
nn Superlinked [m] ar art « [m]
® re
* a . SIE — Superlinked Inference Engine is allan "
i s i
tro :
©) github.com/superlinked/sie 7
] Google DeepMind
ES i a
Slide-Derived Subjects To Review
Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.