Tools AI Minute Newsroom 2026-08-23

A 300-million-parameter model guesses what the big one will say next — and a MacBook answers 2.87 times faster

Liquid AI published DSpark on 20 August: small draft models of roughly 300 million parameters each, paired one-to-one with its LFM2.5 family — the 1.2B Instruct, the 2.6B and the 8B-A1B. The draft proposes a block of nine tokens ahead, the real model checks the whole block in a single forward pass, and anything it disagrees with is discarded. Under greedy decoding the text that comes out is identical to the big model running alone, so benchmark accuracy does not move at all. What moves is time: up to 3.18x on an H100, up to 2.87x on an M4 Max MacBook Pro, the 2.6B model past roughly 140 tokens a second on that laptop, and 57 percent lower latency on average across multi-tool tests. The catch sits in the acceptance rate. On the MATH500 benchmark the 8B-A1B drafter gets 8.27 of every 10 proposed tokens accepted, which is worth 3.18x. On GSM8K it gets 4.02, worth 1.29x — same hardware, same model, roughly a third of the gain.

Why it mattersSpeculative decoding is the rare optimisation with no quality argument attached: the output is provably the same text, so the only open question is how often the small model guesses right. That makes it a useful thing to understand before believing any speedup figure you are sold. A vendor quoting '3x faster' is quoting its best workload; yours decides what you actually get, and the distance between 3.18x and 1.29x here is the whole lesson. It matters most to people running models on their own machine rather than through an API, because this is the kind of change that moves a local model from too slow to bother with to fast enough to leave running.

✓ Verified · 2 sources

Read in the app — free, in 9 languages

A search index built for coding agents rather than people: 70 million READMEs, issues and pull requests, and no key needed to try it
2026-08-23 The standard that plugs AI into your tools was built for a human clicking 'allow'. Its new roadmap admits the caller is now a machine.
2026-08-23 OpenAI has given away the engine under Codex, and its own numbers say the engine is worth more than the model
2026-08-22 A voice model that starts speaking in under 50 milliseconds, and the serving code is open
2026-08-22 Anthropic will let its strongest model hunt flaws in your code — but it will not let you talk to it
2026-08-22

A 300-million-parameter model guesses what the big one will say next — and a MacBook answers 2.87 times faster

Related stories