Fabio Guzman
@FGuzmanAI
Followers
157
Following
1K
Media
22
Statuses
291
On-device ML Engineer | 🤖Passionate about reverse-engineering neural nets | 🚀Optimizing large models for the edge 💻📱
Colombia
Joined May 2011
🚀 Excited to launch CLIP-Finder! 🎉. CLIP-Finder enables semantic searches of images using natural language descriptions and camera input. Built on Apple's MobileCLIP-S0 architecture. Check it out on GitHub: #ComputerVision #CoreML #AppleSilicon
3
25
122
RT @denizbirlikci: 🧵 @karpathy dropped the most compelling lecture at the AI startup school :) Here's my notes summarized with Opus. Main….
0
145
0
Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.
380
665
3K
🧵 AFM UPDATE: Draft Model Discovery! .Found Apple's speculative decoding secret weapon:.Draft Model (48.8M params):. 12 uniform layers, 256 hidden dim.8 attention heads, 768 FFN.Same 153.6K vocab as base model.Just 186MB memory footprint!. The magic: Draft generates 4-8
🚀 Draft technical analysis: 3B Apple Foundation Model (AFMTextV7) .Key architectural findings: .- 3.18B total params (3.11B frozen + 66.6M trainable) .- Two-segment architecture (35+21 = 56 layers) .- Vocabulary: 153.6K tokens, 4K context + RoPE scaling - Native LoRA integration
1
4
11
🚀 Draft technical analysis: 3B Apple Foundation Model (AFMTextV7) .Key architectural findings: .- 3.18B total params (3.11B frozen + 66.6M trainable) .- Two-segment architecture (35+21 = 56 layers) .- Vocabulary: 153.6K tokens, 4K context + RoPE scaling - Native LoRA integration
2
7
54
RT @joefioti: Luminal can discover flash attention entirely automatically. We've been working towards this north star in our search compil….
0
46
0
RT @vercel: Vercel AI Gateway (alpha):. • Built on the @aisdk 5 alpha.• Switch between ~100 AI models without API keys.• Handles auth, usag….
vercel.com
With the AI Gateway, build with any model instantly. No API keys, no configuration, no vendor lock-in.
0
54
0
RT @karpathy: A major mistake I made in my undergrad is that I focused way too much on mathematical lens of computing - computability, deci….
0
1K
0
RT @julien_c: BOOOOM: Today I'm dropping TINY AGENTS. the 50 lines of code Agent in Javascript 🔥. I spent the last few weeks working on thi….
0
68
0
RT @dessatel: Interesting new paper from NVIDIA on "FFN Fusion". They achieved 1.71× faster inference for Llama-3.….
0
1
0
RT @geniegetsme: Sometimes you're just exploring. What if your next favorite spot is right around the corner, and you don't know it. Get th….
0
10
0
RT @geniegetsme: Taking a photo can do more than capture your best memories. It can remind you to share them with your favorite people. It’….
0
60
0
RT @norabelrose: I reached out to the SmolLM2 team asking them to release intermediate training checkpoints, and they just did it! Thanks @….
0
5
0
RT @deepseek_ai: 🚀 Day 1 of #OpenSourceWeek: FlashMLA. Honored to share FlashMLA - our efficient MLA decoding kernel for Hopper GPUs, optim….
0
1K
0
RT @wightmanr: Okay, SigLIP 2 weights for OpenCLIP and timm (image encoder only) are on the @huggingface hub. Merged to main, release proba….
0
13
0
RT @mervenoyann: SigLIP 2 is the most powerful image-text encoder. you can use it to do.> image-to-image search.> text-to-image-search.> im….
0
80
0
RT @osanseviero: SigLIP 2 is out! 👀 Another day, another model:. 🥳Strong vision-language encoder.🔥Outperforms across sizes.📏Dynamic resolut….
huggingface.co
0
41
0
