AI trend breakdowns

Every breakdown we have published, newest first.

The model card's acceptance-length table: GSM8K 5.02 for MTP, 4.36 for DSpark, 5.46 for DFlash 2, down to MT-Bench at 3.74, 3.01 and 4.10, above a definition of acceptance length as completion tokens divided by verification steps.

DFlash 2: a 3.43× drafter that fades to 1.45× under load

Inco AI's block-diffusion drafter reaches 3.43× on GSM8K with one request in flight and 1.45× at concurrency 32. Here is what you pay in memory and variance.

Diagram of the jev-ultrafast decision loop: the page state becomes an indexed element table, one typed request returns an operation and its target, and a small language model is called only for TYPE_TEXT before the executor acts.

Jev Ultrafast: the browser agent that stopped writing text

browser-use's jev-ultrafast finishes a Google Flights search in 7.1 seconds by replacing the agent loop with one typed decision per cycle. Here is the catch.

The laya-coreml README's Measured on M3 Max table: compiled MLX FP16 at 6.94/7.39 ms and 0.4288 J per decision, Core ML ANE FP16 at 4.98/5.31 ms and 0.1540 J, Core ML ANE W8 at 4.88/5.23 ms and 0.1344 J, over 65,598 stable calls.

Laya on the Neural Engine: 4.98 ms and 0.154 J per decision

mizorewww's Core ML port runs one Laya typed decision in 4.98 ms and 0.154 J on an M3 Max. What the parity fixtures prove, and what they do not.

Bar chart and table headed Evaluation on 324 held-out examples: Jev 1.13.0 at 93.21 percent, Bespoke-Nimble-9B at 90.12 percent, Qwen3.8-27B at 84.88 percent, and the untuned Qwen3.5-9B base at 66.36 percent.

Nimble: 2,676 examples to get within three points of Jev

Bespoke Labs fine-tuned Qwen3.5-9B on 2,676 synthetic examples. It reaches 90.12% against Jev's 93.21%, and its evaluation covers six of ten trained domains.