Benchmarks
62 updates on Benchmarks.
LiteRT-LM reports 52 decode tokens per second for Gemma 4 E2B on an Android GPU
Google publishes prefill and decode figures for its on-device runtime, adds Swift and JavaScript APIs, and reports a 2.2x speedup from multi-token prediction.
Prism ML releases Ternary Bonsai, 1.58-bit models with an 8B version at 1.75 GB
Ternary Bonsai 8B needs 1.75 GB, scores 75.5 on Prism ML’s benchmark average and runs at 27 tok/s on an iPhone 17 Pro Max.
Tempo compresses hour-long video with a 2B vision model so a 4B LLM can answer
Meta AI and KAUST report 52.7 on LVBench for a 6B system in which a 2B vision-language model squeezes long video down to about 3 tokens per frame.
Prism ML releases 1-bit Bonsai 8B, an 8.2B-parameter model that needs 1.15 GB
1-bit Bonsai 8B stores all weights in one bit, needs 1.15 GB and runs at about 40 tok/s on an iPhone 17 Pro, according to Prism ML.
iPhone 16 Pro settles 41.5 percent below peak over 20 back-to-back prompts
Four authors ran Qwen 2.5 1.5B 20 times in a row on four edge platforms and report the iPhone 16 Pro settling 41.5 percent below its peak throughput.
Meta's MobileLLM-Flash goes shallow and wide for 1.8 times faster prefill
Meta designed 350M, 650M and 1.4B models by measuring latency on a Galaxy S25, reversing the deep-and-thin rule of the first MobileLLM.
Alibaba adds 0.8B and 2B sizes to Qwen3.5, with 262K context and a vision encoder
Alibaba released Qwen3.5-0.8B and Qwen3.5-2B, dense vision-language models with a 262,144-token context, Apache 2.0 weights and 4-bit MNN builds.
Nemotron-Flash-1B decodes 1.9 times faster than Qwen3-0.6B on an H100
Nvidia designed a hybrid 1B and 3B model family around measured decoding latency on an H100 rather than parameter count, and released three checkpoints.
Intelligence per watt puts local model coverage at 88.7% of real queries
Stanford and Together AI measured 20+ local models on 1M real queries and propose accuracy per watt as the metric for local inference.
lm-Meter times on-device inference and finds prefill, not decode, is the bottleneck
A profiler that runs inside the phone measures each inference phase and reports prefill slowing 158 times as models grow, against 10 times for decode.
Meta trains 140M to 950M reasoning models on 4.2T tokens
MobileLLM-R1 spans 140M to 950M parameters, trained on 4.2T tokens, and Meta scores the 950M model at 74.0 on MATH500.
P/D-Device prefills in the cloud and decodes on the phone, cutting TTFT 60%
Huawei researchers split prefill and decoding between cloud and phone, reporting time to first token down at least 60% and cloud throughput up to 15x.