From the Kynd lab
Notes from the edge
of local AI.
New models, creative engines, and the engineering that makes a private agent team feel fast on the Mac you already own.
LTX 2.5 now fits a 16 GB Mac.
Staged phases, block-streamed weights and tiled decode cut the measured peak from 28 GB to 13.3 GB — with byte-identical output. The 32 GB floor falls in the next update.
Read the engineering note →
Two local leaps, shipping together.
LTX 2.5 Ingredients brings reusable characters to local filmmaking. Qwen 3.8 27B brings a serious new reasoning model to the Kynd agent loop.
Read the launch note →One Mac. A whole agent team.
Inside what we believe is a world-first Mac-local workspace: a main builder, responsive conversation and background specialists, without pretending one prompt is a team.
Read the engineering story →Choose the model your Mac can actually run.
Memory-first setup guides built from the model profiles and measured limits shipping in Kynd.
Make Qwen 3.8 27B faster on an M1 Mac.
Why the loudest new speedup adapters measure slower on M1, M2 and M3—and the prefill tuning that made prompt processing 1.48× faster with identical output.
See the benchmarks → Local LLM · Q2 through Q8Run Qwen 3.8 27B on Mac.
Pick the right quant for 16, 24, 32, 48, 64 or 96 GB—and understand which MLX and GGUF profiles Kynd calibrates today.
Open the quant guide → Local video · 16 GB+Run LTX 2.3 or 2.5 on Mac.
Compare compact and full 2.3 builds with the newest LTX 2.5 int4 path, including honest resolution and Ingredients tiers.
Open the video guide → Local images · 4-bit MLXRun Z-Image Turbo on Mac.
Start from 16 GB with Kynd’s recommended 4-bit pack, or compare the detected-only BF16 route on larger Apple Silicon Macs.
Open the image guide →Read less. Run it locally.
All Build and Workbench features are open for 30 days. No account or card required.