KimiTalk Open-Source Notice

Version: 1.2
Date: 2026-05-26

KimiTalk incorporates open-source software and can prepare local machine learning model weights at the user's request. Current release model downloads are served through KIMIKON-controlled Cloudflare/R2 endpoints (models.kimikon.cc, dmg.kimikon.cc, updates.kimikon.cc), not directly from HuggingFace. HuggingFace remains relevant as an upstream, source, and license reference. Per-component license texts are bundled in Resources/Licenses/ and shown in Settings under Lizenzen & Modellhinweise.

Bundled App Components

• WhisperKit — MIT
• OpenAI Whisper — MIT
• whisper.cpp — MIT
• ggml — MIT
• Sparkle — MIT + external BSD/MIT-style notices
• SpeakerKit (Argmax OSS) — MIT
• pyannote speaker-diarization-community-1 — CC BY 4.0
• FluidAudio (Worker-only runtime for the Parakeet ASR engine) — Apache 2.0
• fastcluster (bundled within FluidAudio, Worker-only, speaker diarization) — BSD 2-Clause (© Daniel Müllner / Google Inc.)
• NVIDIA Parakeet-TDT-0.6B-v3 Core ML weights (FluidInference conversion) — CC BY 4.0
• NVIDIA Parakeet Unified English 0.6B Core ML weights (FluidInference conversion) — NVIDIA Open Model License + CC BY 4.0
• Licensed by NVIDIA Corporation under the NVIDIA Open Model License
• Built with Parakeet Unified English 0.6B by NVIDIA, used under the NVIDIA Open Model License.
• NVIDIA Sortformer Streaming Diarizer v2.1 Core ML weights (FluidInference conversion, Worker-only speaker diarization) — CC BY 4.0
• silero-vad — MIT
• ZIPFoundation — MIT
• Textual 0.5.0 (locally maintained fork) — MIT
• Prism.js 1.29.0 (syntax-highlighting runtime bundled by Textual) — MIT
• swift-concurrency-extras (linked by Textual) — MIT
• swiftui-math (linked by Textual) — MIT
• mlx-swift — MIT
• mlx-swift-lm — MIT
• Swift/Hugging Face packages: swift-transformers, swift-argument-parser, swift-asn1, swift-certificates, swift-collections, swift-crypto, swift-jinja, swift-numerics, swift-syntax, yyjson

Optional Local Speech Helper / Runtime Model Notices

KimiTalk can optionally use separately installed local speech engines and runtime-prepared model weights. The main app keeps visible notices for:

• OmniVoice / Boson Higgs Audio 2
• Meta Llama 3
• Kokoro-82M
• Qwen3-TTS
• Google Gemma 4 / DiffusionGemma
• Qwen local language models
• mlx-community model conversions
• AudioSeal (Local Speech Helper watermark runtime)
• C2PA / c2pa-python (Local Speech Helper provenance runtime)

C2PA is not linked into or vendored by the main KimiTalk.app. Helper-backed AI-audio provenance is created by the separately installed LocalSpeechHelper.app, which bundles c2pa-python/c2pa-rs and carries the full helper NOTICE plus license texts.

Full License Texts

The full Apache License 2.0 text is bundled as Apache-2.0-FULL.txt, and the full Creative Commons Attribution 4.0 International text as CC-BY-4.0-FULL.txt (covers the FluidInference Parakeet conversions, Parakeet-TDT weights, and pyannote community-1). The complete NVIDIA Open Model License for Parakeet Unified English is bundled in Parakeet-Unified-English-LICENSE.txt. Other licenses are bundled in their per-component files. No endorsement by upstream projects is implied unless explicitly stated by those upstream projects.