Research
73 updates on Research.
TinyLLaVA's 3.1B model outscores 7B LLaVA-1.5 on seven of nine benchmarks
Beihang and Tsinghua researchers report a 3.1B vision-language model that beats the 7B LLaVA-1.5 on seven of nine image benchmarks.
MobileVLM V2 runs 1.7B, 3B and 7B vision models on 144 image tokens
Meituan and Zhejiang University report a 1.7B vision language model at 64.2 on six benchmarks and 51.63 tok/s on an NVIDIA Jetson Orin.
TinyLlama pretrains a 1.1B model on 3 trillion tokens
Singapore University of Technology and Design trained a 1.1B model on 3 trillion tokens with 16 A100-40G GPUs and released it under Apache 2.0.
Microsoft releases Phi-2, a 2.7B model it says matches models 25 times larger
The 2.7B base model was trained on 1.4 trillion tokens in 14 days on 96 A100 GPUs, and Microsoft says it matches models up to 25 times larger.
Apple researchers run models twice the size of available DRAM from flash
The LLM in a flash paper loads parameters from flash on demand and reports 4 to 5 times faster CPU and 20 to 25 times faster GPU inference.
Microsoft carries its textbook data recipe from code to reasoning with the 1.3B phi-1.5
The 1.3-billion-parameter model trains on 30B tokens of mostly synthetic data and posts reasoning scores above Llama2-7B in Microsoft evaluations.
MIT HAN Lab publishes TinyChatEngine for 4-bit LLMs on laptops and Raspberry Pi
The from-scratch C and C++ engine from Song Han's group at MIT runs 4-bit LLaMA2-7B on x86, Apple silicon, Raspberry Pi and Nvidia GPUs.
Microsoft trains phi-1 to 50.6 percent on HumanEval with 1.3B parameters
Microsoft Research reports 50.6 percent on HumanEval and 55.5 percent on MBPP for a 1.3B code model trained on under 7B tokens of textbook-quality data.
RWKV trains like a transformer and runs with constant memory per token
The paper releases pretrained RNN weights from 169M to 14B parameters and reports constant time and memory per token during inference.
LaMini-LM distils models from 61M parameters up on 2.58M instructions
MBZUAI distilled 17 instruction models from 61M to 7B parameters on 2.58M synthetic instructions, and reports the 248M model matching LLaMA-7B.
MobileBERT runs in 62 ms on a Pixel 4 with 25.3M parameters
Carnegie Mellon and Google Brain distilled BERT into a 25.3M-parameter model that answers in 62 ms on a Pixel 4 and scores 77.7 on GLUE.
Hugging Face's DistilBERT is 40% smaller than BERT and 71% faster on an iPhone
The 66M-parameter distilled BERT keeps 97% of its teacher on GLUE and weighs 207 MB inside a question-answering app on an iPhone 7 Plus.