## cactus-compute/cactus — v1.14…v2.0

_145 commits._

### Features
- **Add benchmark command and update README with inference speed details** (f91d066)
- **Metal Support (#750)** (eaca207)
- **Enhance dynamic batching support in model component specs** (5780ee2)
- **readme: add Parakeet decode tps numbers (#732)** (a90aae7)
- **Add CQ quantization and platform options to CLI commands** (e4da95d)
- **Add weight revision resolution and version parsing utilities** (6b510c6)
- **Update README to reflect hybrid edge-cloud architecture and add benchmarks** (b85d62a)
- **Add ANSI color support and improve terminal output formatting** (c228a20)
- **Update bundle resolution messages and add bundle management functions** (8e7b71e)
- **Add entitlements file for increased memory limit in iOS tests** (01e6a70)
- **add missing cactus transpile flags** (fd61cbd)
- **Add json string escaping functions and refactor related code for consistency** (09519a0)
- **Add attach_conv_bias method and refactor convolution functions for bias handling** (460eefe)
- **Wire text/image/audio embeddings on the v2 transpiler; add /v1/embeddings (#676)** (add442d)
- **Add v2 local HTTP server (#669)** (da383d3)
- **V2 NPU Support (#664)** (d94478c)
- **Add native chunked prefill for transpiled LLM bundles (#659)** (4e592ff)
- **Add scipy to requirements** (75f9a04)
- **Add fused dense INT4 MLP op for Gemma-4 E2B decode (#617)** (a53f0ce)

### Fixes
- **Fix C++ test build: expose vendored libs include (picojson) (#748)** (f08497a)
- **gemma4: fix fp16 -inf overflow in audio encoder mask, route audio to ANE (#734)** (88c18be)
- **fix(graph): apply SIGMOID/TANH activations in MoE layer (#736)** (4200620)
- **fix slow parakeet loading** (1285315)
- **gemma4: fix 16x audio soft-token magnitude on NPU via single-source scaling (#726)** (f37a9a5)
- **Fix onboarding quickstart bugs and guard destructive `cactus clean` (#722)** (e8c21df)
- **Fix multi-turn chat and tool-call defects across model families (#721)** (852ea07)
- **rag interface fix for lm embeddings (#719)** (7dd1ed5)
- **Fix Gemma 4 NPU vision encoder on non-square images (#718)** (dfd3d8b)
- **fixed needle rope bug, moved repeated rope table instantiations from components into weights (#713)** (5c17824)
- **Fix docs deploy: align SDK handling and rewrite all absolute links** (8a9adf8)
- **fix release workflow** (bed4558)
- **fix clean method** (7a9e1b4)
- **fix and cleanup cli** (4b66a06)
- **bug fixes, cleanups, docs & blog (#657)** (09d0fdc)
- **fix: add missing <iomanip> include for Linux ARM64 builds (#647)** (3227a7f)
- **Fix CQ decode fast path for mmap weights (#633)** (554fd0b)
- **bug fixes** (eac6966)
- **Fix converter logic for models not on cactus huggingface (#608)** (f197e4e)

### Backend
- **v2.0** (3703568)
- **Streamline tests** (3c4d926)
- **Benchmarks (#752)** (76a91a6)
- **Trie-based grammar constraints for Gemma force_tools (#747)** (263c1fc)
- **Cactus code (#745)** (acf3ef2)
- **added cq4 optimizations to cq2 and cq3 (#744)** (2b304a9)
- **Refactor model batch processing and update tests** (ddaf5b1)
- **MOE (#727)** (69e612f)
- **Removed unnecessary null pointer checks (#724)** (e07e26c)
- **asynchronous continupus batching** (ce9bef7)
- **detect SDL2 via pkg-config on all platforms** (712d15f)
- **ragged + continuous batching** (3811cbc)
- **Refactor Cactus Graph for batch inference** (0557094)
- **M3 pro numbers (#742)** (c52d2b2)
- **Package Gemma-4 cloud-handoff probe as a convert asset** (0afa515)
- **squash convert &transpile, and align cli commands (#740)** (0058fd9)
- **coreml dependency** (0d3bff3)
- **parakeet: honor should_stop during TDT transcription (#738)** (7726789)
- **Karen/parakeet probe (#739)** (c4c9cfa)
- **Stt streaming (#730)** (da5649f)
- **Centralize build flags into cactus_flags; shrink libs (#735)** (07d8297)
- **Enhance cloud handoff logging and response handling with detailed reasons for handoff decisions** (76bae59)
- **Simplify CLI: one bundle-resolution ladder + CQ level in list (#715)** (774e1be)
- **fixed lfm and qwen vl handling and run.cpp multiturn behavior (#728)** (eb191b5)
- **Parakeet-TDT NPU acceleration on the Apple Neural Engine (#729)** (09147fc)
- **Refactor entropy handling and update confidence threshold in response JSON** (1daff3b)
- **Enhance command output formatting and update default audio prompt message** (73105da)
- **Update platform options to include 'auto' as default accelerator choice** (8422b4d)
- **Update default model and transcription model IDs for testing** (c413796)
- **convert: fold gemma4 audio projection scaling into the weight reparam (#723)** (a8f634a)
- **LFM2-VL: resolution-native tiling for non-square images** (06946c8)
- **Pad the chunked-prefill tail for sliding-window models (#716)** (d51dbcf)
- **Fuse dispatch for the legacy interleaved CQ4 GEMV (#717)** (4dbc03d)
- **cactus-engine: runtime correctness & memory-safety (#703)** (81857ca)
- **telemetry: JSON-escape records + bound the offline event cache (#704)** (dabdc37)
- **transpiler: JAX attention causality + reduction precision, attention-output layout guard (#705)** (07741ed)
- **transpiler: op routing, fusion & JAX-capture correctness (#706)** (ed1cbcf)
- **convert/CLI: conversion & tokenizer robustness (#707)** (87ad2e6)
- **bindings: JNI & React-Native FFI memory-safety (#708)** (6d0c441)
- **engine+graph: thread-safe last_error + JSON option-parse hardening (#702)** (366d9e2)
- **added language model embed component (#714)** (1db5aad)
- **fixed multiturn gemma tool chat template, fixed qwen and lfm tool chat templates (#711)** (bbb3e9a)
- **Karen/needle v2 (#690)** (7a4ad3a)
- **align chat-template image-token count with vision encoder output** (930d6c8)
- **Use NPU vision encoder for Gemma 4 in chunk-prefill path** (57040c8)
- **Correct docs to match current code (#696)** (14ac567)
- **Long-context: rolling KeyDiff KV-cache compaction + chunked prefill (#686)** (ce1c1b4)
- **Merge cleanup (#692)** (6f3b0c5)
- **Update default model paths and remove unused model references in scripts and tests** (e65fb73)
- **Refactor and optimize kernel functions for improved performance and readability** (55d1cd4)
- **Merge v2 into main** (53f5c50)
- **Jaxtranspilerpath (#681)** (bf31835)
- **Full implmentation into Gemma4 (#668)** (7ab3eac)
- **Propagate test executable failures (#683)** (c4aea8e)
- **Jax Transpiler path (#671)** (c9b08fb)
- **Multiple model fixes (#673)** (2c08602)
- **Make Android scheduling load-aware (#666)** (fb0adae)
- **update docs** (cbc7974)
- **Optimize orthogonal CQ4 LM head layout (#663)** (cad2ebc)
- **Refactor CLI commands and documentation for clarity, update transcribe functionality** (ba15b09)
- **clean cli** (c0f6914)
- **Build system, bindings, CI cleanup, pypi & brew (#661)** (95880ae)
- **Karen/v2 transcribe (#660)** (78719e8)
- **KV Cache Bugfix (reset cache, think token data shift, underflow guard) (#658)** (0e97e13)
- **Cleanup V2 (Cpp Native + CLI Update) (#650)** (52c06ee)
- **Bindings (#653)** (71e109f)
- **Refactor quantization parameters in CLI and publish workflow (#651)** (b2d84f7)
- **Transpiler complete (#652)** (24f7589)
- **Everything (#649)** (2cbcd80)
- **Transpiler complete (#640)** (ecdfafb)
- **Cactus Quant Speedup (#641)** (6e32e47)
- **bugfix** (26094db)
- **bugfix** (2931f7a)
- **bugfix** (ea9d5e8)
- **streamline conversion for CQ** (a085fc7)
- **align c++ official Needle inference (#637)** (7276100)
- **added cactus download cq flag (#631)** (0f87894)
- **Streamline further** (b99db76)
- **cactus-kernels/src/matmul.cpp — Three fixes:** (b7d15d3)
- **Karen/tq v2 debugging (#630)** (60f8721)
- **Pytorch capture -> cactus v2 (#622)** (d36558e)
- **axis reductions fast path -> v2 (#629)** (c40aed0)
- **Hybrid INT8/FP16 decode kernel: 2× faster + sliding-window fast path (#625)** (45753df)
- **inner 1 fast path (#618)** (d917981)
- **optimisations** (20448e2)
- **test fixes** (5669630)
- **optimize tq-kernels** (7a608cd)
- **optimise f16** (dd73244)
- **refactor TQ** (502c751)
- **cleanup engine** (990d7bd)
- **Karen/v2 tq (#615)** (dadeb7e)
- **cactus-graph** (e09b963)
- **cactus-graph** (4951333)
- **Cactus-kernels** (0008507)
- **more refactor** (510752b)
- **Aggressive Refactor!** (b8527fe)
- **Graph remaining ops (#578)** (b338aeb)
- **LFM NPU Fallback + Tokenizer Fixes (#605)** (09d3d79)
- **updated llm.txt** (b5cc03a)
- **mathjax** (d683683)
- **updates** (07f565e)
- **update blog release** (4e72a9f)
- **Update turboquant-h** (dc49d36)
- **TurboQuant-H** (3404637)
- **Gemma4 multimodal multi-turn fixes (#598)** (6f1c63d)
- **Lfm tool arg fixes (#596)** (1fe8a51)

### Tests
- **test: fix test_model.py for v2 transpiled bundle requirement** (828571a)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/cactus-compute/cactus?utm_source=github-action)._