# Arsenal 103 open AI models for your Mac, each its own download, and your agents can use them too. Arsenal is a free, open-source (MIT) menu-bar app for Apple silicon Macs on macOS 26 or newer. It downloads a model's weights from Hugging Face at a pinned revision, runs it on the Mac, and hands it to you and to your agents over MCP and a loopback HTTP API. Each model downloads on its own; all 103 together take 374 GB. Install: ```sh brew install --cask tobynoskillson/tap/arsenal ``` Connect an agent: `claude mcp add arsenal -- arsenal-mcp` ## What the page shows Every example was produced by Arsenal on one MacBook Pro (M5 Max, 128 GB), 21–31 October 2026. Photo and data credits: https://tobynoskillson.com/arsenal/credits.txt - One street photo, four models: RF-DETR Nano finds the motorcycle, MoGe-3 ViT-L gives depth and surface normals, BiRefNet cuts it out. - Better together: FLUX.2 Klein 4B paints a greenhouse from one sentence and MoGe-3 measures its depth and surfaces; TeleOCR reads page one of Moby-Dick (1851), Kokoro 82M speaks the opening and Parakeet v3 Ultra times every word. - See: Grounding DINO Base finds "a man wearing a hat" and SAM 2.1 Hiera Large masks him; BiRefNet lifts a tiger off its ground; Florence-2 Large describes a street. - Fix photos: DDColor colours Mulberry Street, New York, about 1900; RF-DETR finds a boat, EdgeTAM masks it, Big LaMa erases it. - Search by meaning: Ovis VL Embedding 2B finds photos by what they show. - Predict: TimesFM 2.5, t0-alpha, Chronos-Bolt Small and Chronos-2 forecast real series cut at a date, each compared with what was measured afterwards. - All 103: every model in the catalogue by job, with its download size; a lock marks the five gated models. ## Showcases Six jobs done start to finish by Arsenal's own models on one MacBook Pro (M5 Max), 3 November 2026, all on the main page: https://tobynoskillson.com/arsenal/ Each one plays itself on the page: the original, then each model's step in turn. - 1897, with sound and colour: the Lumières' silent film of a train arriving at La Ciotat. Stable Audio 3 makes the locomotive, carriages and crowd (found by Florence-2 and Grounding DINO, checked by CLAP), SeedVR2 3B restores it, RIFE 4.25 doubles the frames, DDColor colours every frame. - A new score for Atlantis: SAM 3.1 tracks the shuttle in NASA's STS-135 launch film, SAM Audio Large lifts out the engines, TIGER DnR keeps the commentary, Parakeet v3 Ultra writes it down, Stable Audio 3 Medium writes a new score. - Mount Vernon, 1963, in colour: SeedVR2 3B restores a U.S. government film, DDColor colours every frame, RIFE 4.25 doubles the frames. - A song taken apart: HTDemucs splits "Alive (Doing It Right)" by Mise (CC BY 4.0) into drums, bass and the rest, Mel-Band RoFormer lifts the voice, Parakeet times every sung word. - Step into a photo: MoGe-3 measures a harbour photo in metres, RF-DETR and Grounding DINO name the boats, boathouse and rocks, SAM 2.1 outlines them; the scene turns in 3D. - From a sentence to an object: FLUX.2 Klein 4B paints a diving helmet from one sentence, BiRefNet lifts it out, TRELLIS.2 4B builds it in 3D. ## Models Sizes are each model's default download. Weights licences are the makers'; any conditions are shown in the app before you download. ### Hear - Audiobox Aesthetics: Rate audio quality. 415 MB. CC-BY-4.0. - CLAP HTSAT Fused: Label sounds from your own words. 618 MB. Apache-2.0. - DiariZen Large v2: Who spoke when, meetings, overlaps. 305 MB. Creative Commons Attribution-NonCommercial 4.0; licence conditions shown before download. - HTDemucs fine-tuned: Separate music into four stems. 336 MB. MIT. - Mel-Band RoFormer Vocals (Kim): Pull voice out of music. 913 MB. MIT. - MOSS-Transcribe-Diarize 0.9B: Transcribe who said what, and when. 1.8 GB. Apache License 2.0. - Nemotron 3 Diarization: Identify who spoke when. 199 MB. OpenMDW-1.1. - Parakeet v3 Ultra: Speech transcription, 25 languages. 1.3 GB. CC-BY-4.0. - Pyannote Community-1: Who spoke when, with overlaps. 33 MB. Creative Commons Attribution 4.0. - Qwen3 ASR 0.6B: Speech transcription, 30 languages. 1.6 GB. Apache-2.0. - Qwen3 ASR 1.7B: Speech transcription, 30 languages. 4.1 GB. Apache-2.0. - SAM Audio Large: Isolate a sound by description. 15.8 GB. SAM License; gated: accept Meta's terms on Hugging Face first. - Smart Turn v3: Detect the end of a spoken turn. 32 MB. BSD-2-Clause. - Streaming Sortformer Diarizer 4spk v2: Who spoke when, live, up to 4. 471 MB. Creative Commons Attribution 4.0. - TIGER DnR: Split dialogue, music and effects. 17 MB. Apache-2.0. - TIGER Speech: Separate two overlapping speakers. 3 MB. Apache-2.0. - Voice Gender Classifier (ECAPA-TDNN): Tell male or female voice. 62 MB. MIT. ### Speak - Fish Audio Drama 3: Expressive speech, voice cloning. 23.8 GB. Fish Audio Community License 1.0; licence conditions shown before download. - Kokoro 82M: Fast English speech. 357 MB. Apache-2.0. - Pocket TTS: Streaming speech, voice cloning. 239 MB. CC-BY-4.0. - Qwen3 TTS 12Hz 1.7B Base: Speech in 10 languages, voice cloning. 4.5 GB. Apache-2.0. - Qwen3 TTS 12Hz 1.7B VoiceDesign: Speech in 10 languages, voice design. 4.5 GB. Apache-2.0. - VoxCPM2: Speech in 30 languages, voice cloning. 5.0 GB. Apache License 2.0. ### Make music and sound - ACE-Step 1.5 Turbo: Songs from lyrics and style. 10.1 GB. MIT. - MiniMax Music 3: Songs from lyrics and a description. 28.5 GB. MiniMax-Music3 Community License; licence conditions shown before download. - Stable Audio 3 Medium: Sound and music generation. 5.2 GB. Stability AI Community; licence conditions shown before download. - Stable Audio 3 Small SFX: Sound effects. 1.7 GB. Stability AI Community; licence conditions shown before download. - YuE2 3B: Songs from lyrics and style. 7.8 GB. CC-BY-NC-4.0; licence conditions shown before download. ### Clean up speech and sound - AudioSR (basic): Restore missing treble at 48 kHz. 6.2 GB. Apache License 2.0. - DeepFilterNet3: Fast speech denoising at 48 kHz. 9 MB. MIT License. - MossFormer2 SE 48K: Speech denoising at 48 kHz. 222 MB. Apache License 2.0. - RE-USE: Speech cleanup. 39 MB. OpenMDW-1.1. - UniPASE: Universal speech enhancement. 2.2 GB. Apache License 2.0. ### Read and translate - Chandra OCR 2: Read handwriting, forms and documents. 10.6 GB. OpenRAIL-M (modified); licence conditions shown before download. - GLM-OCR: Read text from a document. 2.7 GB. MIT. - Index Translate 2B: Translate with glossary and style. 4.6 GB. Apache License 2.0. - Index Translate 9B: Translate with glossary and style. 19.3 GB. Apache License 2.0. - Manga OCR Base: Read Japanese manga text. 444 MB. Apache-2.0. - NuExtract3: Fill JSON templates from documents. 9.1 GB. Apache License 2.0. - TeleOCR: Read text, tables, formulas and code. 2.8 GB. Apache License 2.0. - UVDoc: Flatten a photographed page. 32 MB. Apache-2.0. ### Make and edit images - BiRefNet: Background removal. 444 MB. MIT. - FLUX 3 Image Klein: Image generation and editing. 20.3 GB. FLUX Kommunity License v1.0; gated: accept Black Forest Labs' terms on Hugging Face first. - FLUX.2 Klein 4B: Image generation. 16.0 GB. Apache-2.0. ### See - Depth Anything 3 Metric Large: Metric depth from one photo. 1.3 GB. Apache-2.0. - Depth Anything 3 Small: Relative depth from multiple views. 137 MB. Apache 2.0. - Depth Anything V2 Small: Relative depth from one photo. 99 MB. Apache-2.0. - EdgeTAM On-Device Track Anything Model: Segment an object by click or box. 28 MB. Apache-2.0. - EoMT COCO Panoptic Large 640: Label every object and region. 1.3 GB. MIT. - Florence-2 Base: Caption a photo or read its text. 466 MB. MIT. - Florence-2 Large: Caption a photo or read its text. 1.6 GB. MIT. - Grounding DINO Base: Find objects described in words. 934 MB. Apache-2.0. - MoGe-3 ViT-L: Metric depth and 3D from one image. 1.5 GB. MIT. - OWLv2 Base Patch16 Ensemble: Find objects named in words. 622 MB. Apache-2.0. - RF-DETR Large: Find 80 types of objects. 136 MB. Apache-2.0. - RF-DETR Nano: Find 80 types of objects. 122 MB. Apache-2.0. - SAM 2.1 Hiera Large: Cut out an object by click or box. 898 MB. Apache-2.0. - SAM 2.1 Hiera Tiny: Cut out an object by click or box. 156 MB. Apache-2.0. - SAM 3.1: Select objects and track video. 3.5 GB. SAM License; gated: accept Meta's terms on Hugging Face first. - UI Venus 1.5 2B: Locate screen elements from instructions. 4.9 GB. Apache License 2.0. - V-JEPA 2 ViT-L: Embed video clips. 652 MB. MIT. - ViTPose Base Simple: Find body keypoints. 344 MB. Apache-2.0. ### Fix photos and video - Big LaMa: Photo inpainting. 204 MB. Apache-2.0. - DDColor ModelScope: Photo colorization. 456 MB. Apache-2.0. - FlashVSR v1.1: Video super-resolution. 3.5 GB. Apache-2.0. - GIMM-VFI-R: In-between frames at any moment. 79 MB. S-Lab License 1.0; licence conditions shown before download. - Real-ESRGAN x4plus: Upscale a photo 4x. 67 MB. BSD-3-Clause. - RIFE 4.25: Video frame interpolation. 23 MB. MIT. - SeedVR2 3B: Image and video restoration. 8.4 GB. Apache-2.0. ### 3D - IDOL: One picture to a 3D splat. 6.2 GB. Apache-2.0, with Meta's DINOv3 encoder: gated, accept Meta's terms on Hugging Face first. - TRELLIS.2 4B: Image to textured 3D. 16.2 GB. MIT, with Meta's DINOv3 encoder: gated, accept Meta's terms on Hugging Face first. - TripoSR: Make a 3D model from one photo. 1.7 GB. MIT. ### Search by meaning - astroBERT: Embed astronomy text. 440 MB. MIT. - EmbeddingGemma 2: Search text, images, audio and video. 1.5 GB. Apache-2.0; licence conditions shown before download. - Granite Embedding Small English R2: Embed English text for search. 99 MB. Apache-2.0. - Harrier OSS v1 0.6B: Create text search embeddings. 1.2 GB. MIT. - LCO-Embedding Omni 7B: Search text, images, audio and video. 17.9 GB. Apache-2.0. - Ovis VL Embedding 2B: Find text, images and video by meaning. 4.4 GB. Apache-2.0. - Qwen3 Embedding 0.6B: Multilingual text embeddings. 1.2 GB. apache-2.0. - Qwen3 Reranker 8B: Rank search results by relevance. 16.4 GB. Apache-2.0. - Qwen3-Embedding-8B: Create text search embeddings. 15.1 GB. Apache-2.0. - TIPSv2 L/14: Search photos with text or images. 2.0 GB. Apache-2.0. ### Decide - d1 · 3B: Text and image decisions. 6.3 GB. LFM Open License v1.0; licence conditions shown before download. - elephant-1: Text, image and video decisions. 18.8 GB. MIT. - Laya · English: English decisions. 846 MB. Apache-2.0. - Laya · Multilingual: Decisions in 100+ languages. 678 MB. Apache-2.0. - Laya · Typed decisions: Typed workflow decisions. 846 MB. Apache-2.0. - Mitra-v2 Classifier: Classify table rows in context. 303 MB. Apache-2.0. - Von · 1.1: Option scoring, order-sensitive. 1.6 GB. Apache-2.0. - Von · 1.2: Option scoring, order-independent. 1.6 GB. Apache-2.0. ### Forecast - Chronos-2: Forecast numeric time series. 478 MB. Apache-2.0. - Chronos-Bolt Small: Forecast numeric time series. 191 MB. Apache-2.0. - Granite TimeSeries TTM R3: Forecast numeric time series. 6 MB. Apache-2.0. - T0 Alpha: Forecast numeric time series. 407 MB. Apache-2.0. - T0 Gamma: Forecast series with covariates. 1.0 GB. Apache-2.0. - TimesFM 2.5 200M: Forecast numeric time series. 925 MB. Apache-2.0. - Timestral 209M: Forecast series with covariates. 837 MB. Modified MIT License (Mistral AI); licence conditions shown before download. ### Science - ESM-2 650M: Embed protein sequences. 2.6 GB. MIT. - ESM-2 8M: Embed protein sequences. 31 MB. MIT. - ESM3 Open Small (esm3-sm-open-v1): Protein design and prediction. 5.4 GB. MIT. - Geneformer V2 104M: Embed single cells from genes. 418 MB. Apache-2.0. - MoLFormer-XL Both 10%: Embed molecules from SMILES. 187 MB. Apache-2.0. - ULTRA 50g: Predict missing knowledge-graph links. 1 MB. MIT. Source: https://github.com/TobyNoSkillSon/Arsenal. Sponsor: https://tobynoskillson.com/#support