v2.0

What's Changed

• Lfm tool arg fixes by @justinl66 in #596
• Gemma4 multimodal multi-turn fixes by @ncylich in #598
• LFM NPU Fallback + Tokenizer Fixes by @ParkiratS in #605
• Fix converter logic for models not on cactus huggingface by @cattermelon1234 in #608
• Graph remaining ops by @cattermelon1234 in #578
• Karen/v2 tq by @kar-m in #615
• Add fused dense INT4 MLP op for Gemma-4 E2B decode by @ncylich in #617
• inner 1 fast path by @cattermelon1234 in #618
• Hybrid INT8/FP16 decode kernel: 2× faster + sliding-window fast path by @ncylich in #625
• axis reductions fast path -> v2 by @cattermelon1234 in #629
• Pytorch capture -> cactus v2 by @cattermelon1234 in #622
• Karen/tq v2 debugging by @kar-m in #630
• added cactus download cq flag by @kar-m in #631
• align c++ official Needle inference by @jakmro in #637
• Fix CQ decode fast path for mmap weights by @ncylich in #633
• Cactus Quant Speedup by @justinl66 in #641
• Transpiler complete by @ParkiratS in #640
• fix: add missing include for Linux ARM64 builds by @ramanishka in #647
• Update Cpp pipeline with transpiler, Add Android/IOS Bindings, Fix CLI by @justinl66 in #649
• Transpiler complete by @ParkiratS in #652
• Refactor quantization parameters in CLI and publish workflow by @jakmro in #651
• Bindings by @jakmro in #653
• Cleanup V2 (Cpp Native + CLI Update) by @justinl66 in #650
• KV Cache Bugfix (reset cache, think token data shift, underflow guard) by @justinl66 in #658
• bug fixes, cleanups, docs & blog by @jakmro in #657
• Karen/v2 transcribe by @kar-m in #660
• Build system, bindings, CI cleanup, pypi & brew by @jakmro in #661
• Add native chunked prefill for transpiled LLM bundles by @ncylich in #659
• Optimize orthogonal CQ4 LM head layout by @ncylich in #663
• Make Android scheduling load-aware by @ncylich in #666
• V2 NPU Support by @justinl66 in #664
• Add v2 local HTTP server by @ncylich in #669
• Multiple model fixes by @jakmro in #673
• Jax Transpiler path by @Ayan9074 in #671
• Propagate test executable failures by @psymon-ai in #683
• Wire text/image/audio embeddings on the v2 transpiler; add /v1/embeddings by @ncylich in #676
• Full implmentation into Gemma4 by @ParkiratS in #668
• Jaxtranspilerpath by @Ayan9074 in #681
• V2 by @HenryNdubuaku in #689
• Merge cleanup by @jakmro in #692
• Long-context: rolling KeyDiff KV-cache compaction + chunked prefill by @ncylich in #686
• Correct docs to match current code by @ncylich in #696
• Karen/needle v2 by @kar-m in #690
• Karen/gemmma multiturn template by @kar-m in #711
• fixed needle rope bug, moved repeated rope table instantiations from … by @kar-m in #713
• added language model embed component by @kar-m in #714
• engine+graph: thread-safe last_error + JSON option-parse hardening by @ncylich in #702
• Fix Gemma 4 NPU vision encoder on non-square images by @jakmro in #718
• rag interface fix for lm embeddings by @kar-m in #719
• bindings: JNI & React-Native FFI memory-safety by @ncylich in #708
• convert/CLI: conversion & tokenizer robustness by @ncylich in #707
• transpiler: op routing, fusion & JAX-capture correctness by @ncylich in #706
• transpiler: JAX attention causality + reduction precision, attention-output layout guard by @ncylich in #705
• telemetry: JSON-escape records + bound the offline event cache by @ncylich in #704
• cactus-engine: runtime correctness & memory-safety by @ncylich in #703
• Fuse dispatch for GEMV by @ncylich in #717
• Pad the chunked-prefill tail for sliding-window models by @ncylich in #716
• Fix multi-turn chat and tool-call defects across model families by @ncylich in #721
• Fix onboarding quickstart bugs and guard destructive cactus clean by @rshemet in #722
• convert: fold gemma4 audio projection scaling into the weight reparam by @ncylich in #723
• Parakeet-TDT NPU acceleration on the Apple Neural Engine by @ncylich in #729
• fixed lfm and qwen vl handling and run.cpp multiturn behavior by @kar-m in #728
• Simplify CLI: one bundle-resolution ladder + CQ level in list by @rshemet in #715
• gemma4: fix 16x audio soft-token magnitude on NPU via single-source scaling by @ncylich in #726
• readme: add Parakeet decode tps numbers by @ncylich in #732
• Centralize build flags into cactus_flags; shrink libs by @jakmro in #735
• Stt streaming by @jakmro in #730
• Karen/parakeet probe by @kar-m in #739
• parakeet: honor should_stop during TDT transcription by @rshemet in #738
• squash convert &transpile, and align cli commands by @jakmro in #740
• M3 pro numbers by @ParkiratS in #742
• fix(graph): apply SIGMOID/TANH activations in MoE layer by @yus100 in #736
• gemma4: run the audio encoder correctly and fast on the Apple Neural Engine by @ncylich in #734
• Removed unnecessary null pointer checks by @wasup-yash in #724
• MOE by @ParkiratS in #727
• added cq4 optimizations to cq2 and cq3 by @kar-m in #744
• Cactus code by @jakmro in #745
• Fix C++ test build: expose vendored libs include (picojson) by @rshemet in #748
• Trie-based grammar constraints for Gemma force_tools by @rshemet in #747
• Metal Support by @jakmro in #750
• Benchmarks by @jakmro in #752

New Contributors

• @ramanishka made their first contribution in #647
• @psymon-ai made their first contribution in #683
• @yus100 made their first contribution in #736
• @wasup-yash made their first contribution in #724

Full Changelog: v1.14...v2.0