REPOGEO REPORT · LITE

IST-DASLab/marlin

Default branch master · commit 1f25790b · scanned 6/21/2026, 2:08:07 PM

GitHub: 1,091 stars · 88 forks

Scan history for this repo

Score trend below includes all ready runs (older left, newer right; scroll horizontally if needed). The table is collapsed by default—expand for newest-first rows, 10 per page.

Score trend (left → right: older → newer)

2 ready scans. Expand the table below for newest-first rows (10 per page, paginated).

AI VISIBILITY SCORE

35 /100

Critical

Category recall

0 / 2

Not recommended in any query

Rule findings

1 pass · 1 warn · 0 fail

Objective metadata checks

AI knows your name

3 / 3

Direct prompts that named your repo

HOW TO READ THIS REPORT

Action plan is what to do next — copy-pasteable changes prioritized by impact. Category visibility is the real GEO test: when a user asks an AI a brand-free question that should surface IST-DASLab/marlin, does the AI actually recommend you — or your competitors? Objective checks verify the metadata signals AI engines weight first. Self-mention check detects whether AI even knows you exist by name.

Action plan — copy-paste fixes

3 prioritized changes generated by gemini-2.5-flash. Mark items done after you ship the fix.

OVERALL DIRECTION

highreadme#1

Reposition README opening to emphasize 'kernel' and 'medium batch sizes'

Why:

CURRENT

This is Marlin, a **M**ixed **A**utoR**egressive **Lin**ear kernel (and the name of one of the planet's fastest fish), an extremely optimized FP16xINT4 matmul kernel aimed at LLM inference that can deliver close to ideal (4x) speedups up to batchsizes of 16-32 tokens (in contrast to the 1-2 tokens of prior work with comparable speedup). This makes Marlin well suited for larger-scale serving, speculative decoding or advanced multi-inference schemes such as CoT-Majority.

COPY-PASTE FIX

Marlin is an extremely optimized **FP16xINT4 matmul kernel** specifically designed for **LLM inference at medium batch sizes (16-32 tokens)**. It delivers near-ideal (4x) speedups, significantly outperforming prior work that achieves comparable speedups only at batch sizes of 1-2 tokens. This makes Marlin uniquely suited for larger-scale serving, speculative decoding, or advanced multi-inference schemes like CoT-Majority.

mediumtopics#2

Expand repository topics for better categorization

Why:

CURRENT

4bit, kernel, llm, quantization

COPY-PASTE FIX

4bit, kernel, llm, quantization, fp16, int4, matmul, gpu-acceleration, inference-optimization

mediumhomepage#3

Add a project homepage URL

Why:

COPY-PASTE FIX

Add the official project homepage URL here (e.g., a dedicated documentation site, project page, or relevant research paper link).

Category GEO backends resolved for this scan: google/gemini-2.5-flash, deepseek/deepseek-v4-flash

Category visibility — the real GEO test

Brand-free queries asked to google/gemini-2.5-flash. Did AI recommend you, or someone else?

Same questions for every model — switch tabs to compare answers and rankings.

Recall

0 / 2

0% of queries surface IST-DASLab/marlin

Avg rank

—

Lower is better. #1 = top recommendation.

Share of voice

Of all named tools, what % are you?

Top rival

vLLM

Recommended in 1 of 2 queries

COMPETITOR LEADERBOARD

vLLM · recommended 1×
Hugging Face `transformers` · recommended 1×
`bitsandbytes` · recommended 1×
TensorRT-LLM · recommended 1×
DeepSpeed-MII · recommended 1×

CATEGORY QUERY
How to accelerate LLM inference using 4-bit quantization for medium batch sizes?
you: not recommended
AI recommended (in order):
1. vLLM
2. Hugging Face `transformers`
3. `bitsandbytes`
4. TensorRT-LLM
5. DeepSpeed-MII
6. OpenVINO
7. llama.cpp
AI recommended 7 alternatives but never named IST-DASLab/marlin. This is the gap to close.
Show full AI answer
CATEGORY QUERY
Seeking an optimized kernel for efficient FP16xINT4 matrix multiplication in large language models.
you: not recommended
AI recommended (in order):
1. NVIDIA cuBLASLt
2. NVIDIA CUTLASS (NVIDIA/cutlass)
3. Intel oneMKL
4. TVM (apache/tvm)
5. PyTorch (pytorch/pytorch)
6. TensorFlow (tensorflow/tensorflow)
7. ONNX Runtime (microsoft/onnxruntime)
AI recommended 7 alternatives but never named IST-DASLab/marlin. This is the gap to close.
Show full AI answer

Objective checks

Rule-based audits of metadata signals AI engines weight most.

Metadata completeness
warn
Suggestion:
README presence
pass

Self-mention check

Does AI even know your repo exists when asked about it directly?

Compared to common alternatives in this category, what is the core differentiator of IST-DASLab/marlin?
pass
AI named IST-DASLab/marlin explicitly
AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?
If a team adopts IST-DASLab/marlin in production, what risks or prerequisites should they evaluate first?
pass
AI named IST-DASLab/marlin explicitly
AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?
In one sentence, what problem does the repo IST-DASLab/marlin solve, and who is the primary audience?
pass
AI named IST-DASLab/marlin explicitly
AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?

Embed your GEO score

Drop this badge into the README of IST-DASLab/marlin. It auto-updates whenever the report is rescanned and links back to the latest report — easy public proof that you care about AI discoverability.

Live preview

MARKDOWN (README)

[![RepoGEO](https://repogeo.com/badge/IST-DASLab/marlin.svg)](https://repogeo.com/en/r/IST-DASLab/marlin)

HTML

<a href="https://repogeo.com/en/r/IST-DASLab/marlin"><img src="https://repogeo.com/badge/IST-DASLab/marlin.svg" alt="RepoGEO" /></a>

Pro

Subscribe to Pro for deep diagnoses

IST-DASLab/marlin — Lite scans stay free; this card itemizes Pro deep limits vs Lite.

Deep reports10 / month
Brand-free category queries5 vs 2 in Lite
Prioritized action items8 vs 3 in Lite