v2.1.0

What's Changed

• docs: fix "HuggigFace" typo in README by @DaveyEke in #746
• M3 pro bench results by @ParkiratS in #762
• add wider cq support by @jakmro in #763
• add needle support by @jakmro in #765
• add tool calling support to run env by @jakmro in #766
• optimize attention and gemv kernels by @jakmro in #776
• Custom transpiler by @ParkiratS in #778
• Include directly for std::setw/setfill/setprecision (fixes build on libstdc++/Pi) by @nitishagar in #782
• Fix per-matrix CQ4 reciprocals in Metal batch transform by @clumsylad21 in #784
• Bucket the Parakeet audio encoder by @jakmro in #791

New Contributors

• @DaveyEke made their first contribution in #746
• @nitishagar made their first contribution in #782
• @clumsylad21 made their first contribution in #784

Full Changelog: v2.0.1...v2.1.0

v2.0.1

What's Changed

• Remove coreml by @jakmro in #754
• fix brew by @jakmro in #755
• fix brew by @jakmro in #758

Full Changelog: v2.0...v2.0.1

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

v1.14

What's Changed

• Enable SDL2 live mic recording in chat build by @rshemet in #579
• fix gemma4 conversion and cli by @jakmro in #581
• Made non-thinking default for gemma4 & Introduce simultaneous multimodalitiy by @ParkiratS in #582
• gemma4 vision memory optimizations by @jakmro in #590
• added a prompt fix for better vision+audio prompt, removed code that … by @kar-m in #589
• added default confidence and wired that for vlm by @kar-m in #591
• Gemma4 fixes by @ParkiratS in #588
• Gemma4 tool calling and RoPE fixes by @ncylich in #594
• increased rolling window size for gemma4 by @kar-m in #593

Full Changelog: v1.13...v1.14

v1.13

What's Changed

• Improve model publishing error handling in workflow by @jakmro in #546
• Fix VLM prefill cache reuse image path slicing by @KayaanT in #545
• Add cache cleanup for Hugging Face models in export_and_publish_model by @jakmro in #547
• Add branch override option to publish workflow by @ncylich in #548
• Fix Gemma 3n conversion memory usage by @ncylich in #549
• Update blog URL in README.md by @amerkld in #544
• Parakeet streaming fix by @ParkiratS in #551
• Tool call prompt formatting by @jakmro in #558
• Feature Cleanup by @ParkiratS in #559
• Add Whisper v3 (large-v3) support by @ncylich in #557
• Add LFM2.5 VL 450M by @yujonglee in #567
• LFM2-VL-450M: fix garbled output (vision tower, template, kernel) by @ncylich in #565
• Change default transcribe model to parakeet-tdt-0.6b-v3 by @rshemet in #566
• Expose min_p and repetition_penalty in completion options by @DuFanYin in #560
• Graph save load by @cattermelon1234 in #556
• Pyannote features and optimizations by @jakmro in #571
• Karen/needle by @kar-m in #574
• fix apple i8mm detection to use runtime sysctl check by @DuFanYin in #562
• Fix streaming transcribe by @jakmro in #576

New Contributors

• @amerkld made their first contribution in #544
• @DuFanYin made their first contribution in #560

Full Changelog: v1.12...v1.13

v1.12

What's Changed

• Gemma 4
• Versioned docs with quickstart and SDK chooser by @rshemet in #522
• add CACTUS_CLOUD_API_BASE by @yujonglee in #521
• Model weights discoverability by @jakmro in #520
• fix: set telemetry framework to "rust" for Rust bindings by @rshemet in #519
• fix: prefer local libcactus over system-installed in test build by @rshemet in #516
• Engine updates: tool calling, compute_entropy, converter improvements by @ncylich in #517
• cactus_prefill by @mhayes853 in #512
• Missing torch ops by @cattermelon1234 in #518
• Docs fixes by @jakmro in #524
• docs: document custom vocabulary support for transcription by @ayushmk7 in #525
• Youtu by @jakmro in #530
• Clean up Hugging Face cache after model export by @jakmro in #533
• Added parakeet optimizations for apple by @ParkiratS in #534
• Cactus torch api clean by @cattermelon1234 in #529
• add CACTUS_CLOUD_HEADERS support by @yujonglee in #531
• Parakeet encoder optimization by @ParkiratS in #535
• Update docs site_url to docs.cactuscompute.com by @rshemet in #527
• Custom vocabulary support for Parakeet TDT by @rshemet in #532
• Fix VLM crash when adding multiple images in multi-turn conversation by @FarooqMulla in #539
• New LayerNorm Kernel by @nshejwalkar in #540
• Add pyannote/segmentation-3.0 speaker diarization (10ms, 976× realtime) by @rshemet in #538
• Tinyllama by @ParkiratS in #536

New Contributors

• @cattermelon1234 made their first contribution in #518
• @ayushmk7 made their first contribution in #525
• @FarooqMulla made their first contribution in #539

Full Changelog: v1.11...v1.12

v1.11

What's Changed

• Fix/issue#490 by @lennartvoelz in #491
• simplify and align sdks by @jakmro in #489
• remove models by @jakmro in #492
• Update model configurations and enhance workflow settings in publish_… by @jakmro in #495
• Update workflow to use macos-latest instead of macos-latest-xlarge by @jakmro in #496
• Add dynamic max_tokens estimation based on audio length in cactus_tra… by @jakmro in #499
• macOS: link clang_rt.osx to fix SME2 (_arm_tpidr2*) link failures under rustc by @yujonglee in #498
• Add FFI log control: cactus_log_set_level and cactus_log_set_callback by @yujonglee in #497
• Karen/qwen3p5 by @kar-m in #481
• CLI upgrades by @rshemet in #504
• feat(stt): custom vocabulary biasing for all speech models by @vyomshah05 in #451
• Add Gemma 3N (text-only) model support by @ncylich in #493
• fix: make FunctionGemma prompt formatting strict by @lennartvoelz in #502
• fix: apply logit bias before greedy sampling by @ncylich in #507
• remove redundant file linking for tie_word_embeddings by @jakmro in #506
• Port general engine improvements for TinyLlama by @ncylich in #513
• Speech-to-Text Timestamps by @jakmro in #515

New Contributors

• @lennartvoelz made their first contribution in #491

Full Changelog: v1.10...v1.11

v1.10

What's Changed

• Enhance model publishing workflow with detailed metadata and licenses by @jakmro in #459
• Added parakeet to publish to hf yaml by @ParkiratS in #464
• Update telemetry for supported platforms by @justinl66 in #465
• added back moe weight conversion by @kar-m in #468
• adjust manual workflow for model publish by @jakmro in #470
• Parakeet blog by @ammesatyajit in #467
• perf: add FP16 fast path for LayerNorm by @yujonglee in #433
• Issue #406: Bilinear + Depthwise Optimizations by @PiyawanChaiprasit2006 in #466
• ARM SME2: Accelerate MatMul FP16 by @aarav18 in #457
• build: add Objective-C ARC support for NPU sources by @jakmro in #475
• long transcription by @jakmro in #482
• Language detection by @ParkiratS in #471
• Parakeet tdt by @ParkiratS in #476
• kotlin: expose forceTools in CompletionOptions by @rshemet in #484
• Update model list in README and publish_to_hf.yml with new LiquidAI m… by @jakmro in #487
• test: updated rag test conditions by @nshejwalkar in #488
• optimize scale correction in cactus_attention_f16_h64 by @jakmro in #485
• fix greedy sampler ignoring logit suppression by @jakmro in #486

New Contributors

• @PiyawanChaiprasit2006 made their first contribution in #466
• @aarav18 made their first contribution in #457

Full Changelog: v1.9...v1.10

v1.9

Whats New

1. 50% faster int4
2. Parakeet models
3. LFM2-MOE models
4. BugFixes
5. Hybrid Inference

PRs

• fix stt test and add cpp ci by @yujonglee in #413
• add IRFFT by @yujonglee in #425
• fixed lfm2 vlm lmhead issue that came in with hf 5.0.0 by @kar-m in #426
• raspberry pi numebrs and linux fixes by @kar-m in #437
• Added parakeet model by @ParkiratS in #443
• Adding parakeet graph by @ParkiratS in #446
• Parakeet kernel by @ParkiratS in #445
• added cloud fallback and documentation+tests by @kar-m in #369
• Parakeet FFI by @ParkiratS in #447
• Parakeet convert and tests by @ParkiratS in #444
• Hybrid transcription blog post by @rshemet in #449
• Fixed missing engine changes by @ParkiratS in #453
• feat(python): add context manager support for safe resource cleanup by @yogyam in #412
• Completed ubuntu CICD pipeline by @ncylich in #455
• Tie-embed-conversion-fix by @ncylich in #454
• tiny graph fix and added benchmark by @kar-m in #456

Full Changelog: v1.8...v1.9

Breaking changes

Weights unfortunately need to be refreshed for this :(

v1.8

What's Changed

• Kernel optimisations by @HenryNdubuaku in #397
• Improve INT4 by @ncylich and @jrajala6 in #343
• add einops dependency to requirements by @jakmro in #371
• Add language parameter support for Whisper transcription by @rshemet in #384
• added moe support for lfm by @kar-m in #374
• Add raw FFI binding for Rust by @yujonglee in #382
• fix: handle spaces in paths when running shell commands by @adithya-n05 in #377
• fixing sentencepiece detection for transformers 5.0+ (still backwards compatible) by @ncylich in #373
• Improve Telemetry by @mhayes853 in #372
• proprietry commit by @HenryNdubuaku
• Update performance metrics for iPhone 13 Mini and Galaxy A56 by @jakmro in #386
• fix: improve version sorting and enhance model export tagging by @jakmro in #387
• Add Rust SDK and language parameter documentation by @rshemet in #389
• Basic addition of int4 functionality by @jrajala6 in #343
• add scalar log by @yujonglee in #390
• fix assertion and linux build in rust test by @yujonglee in #392
• Justin/api fixes by @justinl66 in #380
• Update telemetry by @justinl66 in #394
• docs: add compatibility guidelines for runtime and weights by @jakmro in #398
• add STFT_COMPLEX, derive stft_magnitude via graph composition by @yujonglee in #395

New Contributors

• @yujonglee made their first contribution in #382
• @adithya-n05 made their first contribution in #377

Full Changelog: v1.7...v1.8

Note:
This breaks the weights.