mac retweeted
✴️ BENCHMARKED: CLOUDFLARE WITHOUT THE CLOUD
@Cloudflare's new open-weight document reader, Clairic-flash 9B, running fully local on my 64GB M5 Pro Mac mini.
As promised, here's tonight's benchmark.
- It correctly extracted all 34/34 pieces of information from an image of a receipt, a pdf of an invoice and a form
- I ported it to MLX using @SpaceXSI and it's blazing fast on Apple silicon
- 40.5 tok/s at 4-bit (3.2x faster than with PyTorch!)
- 36 tok/s even with a long context size (16K)
- Model loading time: 3.2 s
- Memory usage: 7.2 GB
- 33/34 fields from the same documents' text on MLX
A local document reader LLM that actually works.
Huge release from Cloudflare 🧡
JUST IN: @Cloudflare is a neolab now and it's shipping open weights under Apache 2.0
Clairic-flash is the 9B, fine-tuned from Qwen3.5-9B. The BF16 weights take up about 20GB and I was able to run inference on my 64GB Mac with room to spare.
Every extracted field also comes with a confidence score, which will come in handy (esp. for receipts and invoices).
Will post benchmark results tonight.
mac retweeted
"Women also who administer drugs to cause abortion, as well as those who take poisons to destroy unborn children, are murderesses" (Letter 188).
+ St. Basil the Great
Taylor Swift: “I’m a Christian” and People With Real “Christian Values” Support Abortion lifenews.com/2020/02/04/tayl…
mac retweeted
Replying to @HowToPrompt__
AtlasKV is cool, but “1B triples in 20GB VRAM” is misleading.
20GB is just the inference-time GPU footprint after hierarchical pruning. The full index still lives on host memory or disk (often 1–10+ TB).
You’re not stuffing a billion triples into the model. You’re doing approximate lookups over a massive external store that never fully touches the GPU.
Think of AtlasKV as a fast, lossy thumbnail like the profile pic on this post. It uses far less context than RAG, but it gives up verbatim fidelity and fine-grained factual detail.
RAG still pulls the full-resolution source: exact passages, citations, timestamps, and permissions.
AtlasKV is a compressed memory layer that works when approximate grounding is good enough and you control the model weights.
The advantages of RAG are decisive in domains such as mine (healthcare). A thumbnail of a patient chart isn’t the chart. Same for any use case that needs exact recall, auditability, access controls, or instant doc sync.
Smart parametric memory technique for open models. Not a RAG killer.
mac retweeted
“You have nothing in common with your ancestors”
The Saxons called stuff that they didn’t like “gay” until their word for “gay” (bæddeling) became our word “bad”. They were exactly like me
mac retweeted
>wait 16 years for GTA 6
>cant use guns in gay nightclubs
>cant runover protected minority groups
>GTA gf is a fat brown latina torta
>cant cheat on said fat brown latina gf
>no hookers or cocaine
>first 6 hours of the game involve doing doing chores for nagging bitch girlfriend brown torta
>all dialogue is written by feminist girlboss libtards who hate men
>male main character simps for a girl that ghosted on him for a year
>indian devs have left 6 million coding errors such as zero plane crashes
>combat system glitches and sucks
>barely does 20 fps, (gta 5 did 60)
>basically paid 100 usd to be a simp cuck to a brown gf for 200 hours
>realize actual fun video games and the non gay world you grew up are never coming back
>nigga
>its over
mac retweeted
Just heard one of the funniest lines of the year:
"I want so many deportations the Irish start to get nervous."
Absolute cinema.
mac retweeted
*NSA IT DISRUPTION LINKED TO RAPID PATCHING OF AI-DISCOVERED PHP FLAWS: CNN
*DNI CLAYTON CITES 'VERY HIGH CONFIDENCE' RECENT OUTAGES NOT FROM HACKING

































