TensorFold 1.1.9 is out 🙌 A few results against 1.1.8 from the release tests: - Qwen3.8-27B on GB10 with an MLX 4-bit checkpoint: 9% to 11% faster prefill. - Kolibri-1 on GB10: 3% to 8% faster prefill. - Qwen3.8-27B on M5 Ultra: 3.7% to 4.5% faster prefill. - Flash Next on M3 Ultra: sampled prose went from 122.5 to 136 tok/s. Copy-heavy edits ran between 212 and 257 tok/s. Every model in the release comparison matched 1.1.8 token for token. There's also better follow-up reuse for Qwen3.8-27B on Macs with limited memory, opt-in Qwen3.5-2B conversation caching, Nemotron and GLM improvements, clearer drafter reporting and a fix for automatic compaction on small context windows. Huge thanks to everyone who helped with this release: @Bizuayeu @jishvenugopal @grantoverton And on GitHub: jschmied, yguanrong, akol1, RS91Dev, Deesha08, laudney and ppardus. To everyone contributing, testing, sharing hardware access and supporting TensorFold, thank you. Apologies if I've missed anyone. @volatilemarkts @plotarmordev @petruspennanen @MiaAI_lab