good reminder: I need to check my llama.cpp quants I suspect I’m leaving perf on the table.
@reach_vb
-
Llava o1: Open-Source Vision Language Model with CoT
By
–
Llava o1: https://
huggingface.co/Xkev/Llama-3.2
V-11B-cot
… -
Major Open Source LLM Releases: Pixtral, Tülu Compete With Claude
By
–
Massive week for Open AI/ ML: @MistralAI Pixtral & Instruct Large – ~123B, 128K context, multilingual, json + function calling & open weights @allen_ai Tülu 70B & 8B – competive with claude 3.5 haiku, beats all major open models like llama 3.1 70B, qwen 2.5 and nemotron Llava
-
Open Weights and Science Drive AI Use Cases Forward
By
–
Many such use-cases being powered by open weights & science – so many low hanging fruits! https://t.co/yJMgujXcrG
— Vaibhav (VB) Srivastav (@reach_vb) 24 novembre 2024Many such use-cases being powered by open weights & science – so many low hanging fruits!
-
Apple Releases Fast CoreML Models for iPhone Performance
By
–
Pretty cool! @Apple released blazingly fast CoreML models AND an iOS app to run them on iPhone! ⚡
— Vaibhav (VB) Srivastav (@reach_vb) 23 novembre 2024
> S0 matches OpenAI's ViT-B/16 in zero-shot performance but is 4.8x faster and 2.8x smaller
> S2 outperforms SigLIP's ViT-B/16 in zero-shot accuracy, being 2.3x faster, 2.1x… pic.twitter.com/p9hPoajOtvPretty cool! @Apple released blazingly fast CoreML models AND an iOS app to run them on iPhone! > S0 matches OpenAI's ViT-B/16 in zero-shot performance but is 4.8x faster and 2.8x smaller > S2 outperforms SigLIP's ViT-B/16 in zero-shot accuracy, being 2.3x faster, 2.1x
-
Apple Releases AIMv2 Vision Encoders Outperforming CLIP
By
–
New open release from @Apple – AIMv2 – large scale vision encoders > Outperforms CLIP and SigLIP on major multimodal understanding benchmarks
> Beats DINOv2 on open-vocabulary object detection and referring expression comprehension
> Strong recognition performance w/ -
Bfloat16 vs Quantization: Performance Trade-offs in Model Deployment
By
–
Bfloat16 or nothing! FWIW – all the models deployed on Hugging Chat are bf16. Quants are good for local/ hobby use – however you always leave perf on the table.