Liquid AI's DSpark: 3.1x faster local vision inference, open weights
A 280M drafter speeds LFM2.5-VL-3B up to 3.13x on an M5 Max, greedy output unchanged. Also: Nvidia open-sources SoL-Pi, cutting coding-agent tokens ~49%.

Copy markdown
DSpark: a 280M drafter, up to 3.13x faster VLM decoding
Liquid AI's LFM2.5-VL-3B-DSpark is a 279.5M-param speculative-decoding drafter that proposes tokens the base 3B vision model verifies in a single pass: 2.30-3.13x faster decode on an M5 Max (MLX-VLM), 1.57-2.14x on an M3 Ultra (llama.cpp), and 2.04-2.66x on an H100 (SGLang), with greedy output identical to the base model. Weights ship on Hugging Face in Safetensors and GGUF, so you can bolt faster local multimodal inference onto a laptop this weekend; the LFM Open License allows free commercial use under $10M revenue.
Nvidia open-sources SoL-Pi: ~49% fewer coding-agent tokens
SoL-Pi is an auto-researched agent harness that trims token use 45-49% versus Nvidia's Pi harness and 35-64% versus native model harnesses, roughly a third off your list-price API bill, with minimal quality loss on GPT-5.6 Sol, Claude Opus 5, and Codex. Its four surviving tricks (fusing edits into one tool call, compacting context at subtask boundaries, caching repeated tool outputs, and offloading log-reading to cheaper agents) beat 148 other candidates across 3,000+ runs. The code is on GitHub to drop into your own harness.