LFM2.5-DSpark: up to 3.2x faster inference, H100 down to MacBook
Liquid AI's draft models add speculative decoding to LFM2.5 at zero quality cost — GGUF out now. Plus GitHub Copilot's agent lands in Microsoft Teams.
Copy markdown
One draft model, up to 3.18x the tokens
LFM2.5-DSpark bolts a ~300M-param draft model onto each LFM2.5 target for speculative decoding: the 8B-A1B hits 3.18x on MATH500 (428→1,362 tok/s) and the 2.6B hits 2.87x on MT-Bench. Output is identical to greedy decoding — pure speed, no quality trade.
It runs on your MacBook, not just an H100
GGUF builds ship day-one for llama.cpp and SGLang. On an M4 Max the 1.2B sees up to 2.87x on HumanEval (136→389 tok/s) and function-calling latency drops ~57% — fast enough to make on-device agent loops practical.
Elsewhere: Copilot's agent lands in Microsoft Teams
Mention @GitHub in a Teams thread to spin up a cloud-sandbox coding agent that works async and opens a PR — the Teams counterpart to GitHub's new Slack integration. Needs a paid Copilot plan; cloud sessions bill separately via AI credits.