Nextended.Aspire.Hosting.LocalAI — API reference

🇩🇪 Diese Seite auf Deutsch

The complete public surface of Nextended.Aspire.Hosting.LocalAI, generated from the compiled assembly.

Generated

This page is generated by tools/ApiRef from the compiled assembly — it includes members with no XML comment and therefore cannot drift from the code. Do not edit by hand.

↩ Back to the package page

Nextended.Aspire.Hosting.LocalAI

AceStepApiResource

class

The ACE-Step 1.5 server (Gradio + REST API) backing an AceStepUiResource.

Constructors

• AceStepApiResource(string name)
  The ACE-Step 1.5 server (Gradio + REST API) backing an AceStepUiResource.

AceStepUiBuilderExtensions

static class

Adds the ace-step-ui music studio (a local Suno-style UI: song library, lyrics editor, stem separation, audio editor) to the stack.

AceStepUiOptions

class

Options for WithAceStepUi. Defaults run the official ACE-Step 1.5 server image plus the ace-step-ui studio built from source (the repo ships no container image).

Constructors

• AceStepUiOptions()

Properties

• ApiEnvironment : IDictionary<string, string> { get; }
  Extra environment variables for the ACE-Step server container (applied last).
• ApiGitRef : string { get; set; }
  Git branch/tag of ApiRepository to build (source build only). Default v0.1.4 — the revision whose /generation_wrapper signature the UI is built against. If you change it, pin UiGitRef to a UI revision matching that server (see BuildApiFromSource).
• ApiHostPort : int? { get; set; }
  Fixed host port for the ACE-Step server (random if null). The UI talks to it internally either way; the endpoint also serves ACE-Step's own Gradio UI, handy for debugging.
• ApiImage : string { get; set; }
  ACE-Step server image (without tag) — only used when BuildApiFromSource is false. Default ghcr.io/ace-step/ace-step-1.5.
• ApiRepository : string { get; set; }
  Git repository the ACE-Step server is built from (source build only). Default https://github.com/ace-step/ACE-Step-1.5.
• ApiTag : string { get; set; }
  ACE-Step server image tag (see ApiImage). Default latest. WARNING: released images (0.1.8/latest) do NOT match the current UI's positional Gradio signature — only use a prebuilt image together with a UiGitRef built against exactly that server version.
• BuildApiFromSource : bool { get; set; }
  Build the ACE-Step server from ApiRepository@ApiGitRef via a generated Dockerfile instead of pulling ApiImage:ApiTag. Default true, and deliberately so: ace-step-ui calls the Gradio /generation_wrapper with POSITIONAL arguments, so server and UI must pair exactly. The UI's argument list matches the ACE-Step v0.1.4 signature; newer releases inserted parameters, shifting positions (symptom: "Value: is not in the list of choices ['euler','heun']") — and GHCR offers no v0.1.4 image, hence the source build. Set false only with a matching prebuilt image (see ApiTag).
• ConfigPath : string { get; set; }
  ACE-Step DiT model config (ACESTEP_CONFIG_PATH). Default acestep-v15-turbo; null/empty = server default.
• Environment : IDictionary<string, string> { get; }
  Extra environment variables for the UI container (applied last — overrides the built-in wiring).
• LmModelPath : string { get; set; }
  ACE-Step language model (ACESTEP_LM_MODEL_PATH) powering the UI's "thinking"/enhance features. Default acestep-5Hz-lm-4B (~8 GB extra VRAM); set null/empty to skip loading a LM.
• NodeImage : string { get; set; }
  Base image for the generated UI Dockerfile. Default node:22-bookworm.
• PexelsApiKey : string { get; set; }
  Optional Pexels API key for the UI's video-background feature (PEXELS_API_KEY).
• UiGitRef : string { get; set; }
  Git branch/tag of UiRepository to build. Default main — pin a tag/commit-ish branch for reproducible builds.
• UiRepository : string { get; set; }
  Git repository the UI is built from. Default https://github.com/fspecii/ace-step-ui.

AceStepUiResource

class

The ace-step-ui music studio (Suno-style frontend for ACE-Step), built from source.

Constructors

• AceStepUiResource(string name)
  The ace-step-ui music studio (Suno-style frontend for ACE-Step), built from source.

ImageModel

class

A text-to-image model reference. Implicitly convertible from String (any LocalAI gallery name, OCI/huggingface URI or config URL) and from KnownImageModel for a friendly, discoverable API: imagegen.AddModel(KnownImageModel.Flux1Schnell); imagegen.AddModel("dreamshaper");

Constructors

• ImageModel(KnownImageModel model)
• ImageModel(string name)

Methods

• F16Of(KnownHuggingFaceImageModel model) : bool
  Whether to load the fp16 file variant. This is per-repo: some repos ship ONLY *.fp16.safetensors (need f16=true), others ONLY default-named weights (need f16=false); the wrong setting fails with "variant=fp16, no such files" or "necessary safetensors weights ... (variant=None)". Values verified against each repo's unet folder. "BOTH"-repos use fp16 to save VRAM.
• NameOf(KnownImageModel model) : string
  Resolves the gallery name of a KnownImageModel via its DescriptionAttribute.
• RepoOf(KnownHuggingFaceImageModel model) : string
  Resolves the HF repo id of a KnownHuggingFaceImageModel.
• StepsOf(KnownHuggingFaceImageModel model) : int
  Recommended sampler steps for a known HF model (turbo models need few).

Properties

• Name : string { get; }
  The model id consumers pass to the API (and shown in /v1/models).

KnownEmbeddingModel

enum

Well-known text-embedding models from the LocalAI gallery, installed via AddEmbeddingModel and served on /v1/embeddings (semantic search / RAG). Any other works via the string overload.

Values

• AllMiniLmL6v2
  all-MiniLM-L6-v2 — sehr kleiner, schneller Sentence-Transformer-Klassiker.
• AllMiniLmL6v2OpenVino
  all-MiniLM-L6-v2 (OpenVINO) — CPU-optimierte Variante.
• BertEmbeddings
  BERT embeddings — klein, universell, guter Default.
• BgeM3
  BGE-M3 (ColBERT) — multilingual, Multi-Vektor.
• EmbeddingGemma300m
  Google EmbeddingGemma 300M.
• GraniteEmbeddingEn
  IBM Granite Embedding 125M (Englisch).
• GraniteEmbeddingMulti
  IBM Granite Embedding 107M (multilingual).
• MultilingualE5Base
  Multilingual E5 Base (OpenVINO) — mehrsprachige Embeddings, CPU-optimiert.
• NomicEmbedText
  Nomic Embed Text v1.5 — starke, offene Embeddings.
• Qwen3Embedding0_6b
  Qwen3 Embedding 0.6B.
• Qwen3Embedding4b
  Qwen3 Embedding 4B.
• Qwen3Embedding8b
  Qwen3 Embedding 8B.
• Qwen3VlEmbedding2b
  Qwen3-VL Embedding 2B — multimodale (Text+Bild) Embeddings.
• Qwen3VlEmbedding8b
  Qwen3-VL Embedding 8B — größere multimodale Embeddings.
• Qwen3VlReranker2b
  Qwen3-VL Reranker 2B — multimodaler Reranker für RAG.
• Qwen3VlReranker8b
  Qwen3-VL Reranker 8B — größerer multimodaler Reranker.
• value__

KnownHuggingFaceImageLora

enum

Curated, well-known text-to-image LoRA adapters hosted on HuggingFace. A LoRA is NOT a standalone model — it is applied on top of a base checkpoint. AddModel/AddImageLora generate a diffusers config that (a) downloads the LoRA .safetensors into the model dir and (b) applies it on the matching base (here all on SDXL base 1.0). The DescriptionAttribute holds the LoRA repo id; base + weight file are resolved in ImageLora. These are general-purpose style LoRAs — put your own domain-specific LoRAs directly in your AppHost via AddImageLora(name, base, "owner/repo"). Most need a trigger phrase in the prompt (noted below).

Values

• GraphicNovel
  Graphic Novel Illustration (blink7630) — inked comic/graphic-novel look. Trigger: "graphic novel illustration".
• IkeaInstructions
  IKEA Instructions (ostris) — flat IKEA-manual illustration style. Trigger: "ikea instructions".
• LineArtManga
  Line-Art / Manga (artificialguybr, LineAniRedmond V2) — clean line art. Trigger: "LineAniAF" / "lineart".
• Papercut
  Papercut (TheLastBen) — layered paper-cutout look. Trigger: "papercut".
• PixelArt
  Pixel Art XL (nerijs) — retro pixel-art style. Trigger: "pixel art".
• ThreeDRender
  3D Render Style (goofyai) — clean 3D-render aesthetic. Trigger: "3d style" / "3d render".
• ToyFace
  Toy Face (CiroN2022) — cute 3D toy/figurine faces. Trigger: "toy_face".
• value__

KnownHuggingFaceImageModel

enum

Curated text-to-image models hosted on HuggingFace (diffusers format) that are NOT in the LocalAI gallery. Loaded via AddHuggingFaceModel, which generates a diffusers model config on the fly. The DescriptionAttribute holds the HF repo id. All repos verified public/ungated.

Values

• AbsoluteReality
  AbsoluteReality v1.8.1 (SD1.5) — photoreal SD1.5 all-rounder.
• AnimagineXL31
  Animagine XL 3.1 — high-quality anime SDXL (predecessor of 4.0).
• AnimagineXL4
  Animagine XL 4.0 — high-quality anime SDXL.
• CounterfeitV3
  Counterfeit V3.0 (SD1.5) — high-quality anime SD1.5.
• CyberRealistic
  CyberRealistic v3.3 (SD1.5) — photoreal SD1.5, renders NSFW.
• DeliberateV2
  Deliberate v2 (SD1.5) — versatile, widely used SD1.5 model.
• DreamShaper7
  DreamShaper 7 (SD1.5) — popular all-rounder, predecessor of DreamShaper 8.
• DreamShaper8
  DreamShaper 8 — extremely popular SD1.5 all-rounder.
• DreamShaperXLTurbo
  DreamShaper XL v2 Turbo — fast SDXL turbo (few steps).
• EpicRealism15
  epiCRealism (SD1.5) — beloved photoreal SD1.5 model.
• EpicRealismXL
  epiCRealism XL v7 (Final Destination) — photorealistic SDXL.
• IllustriousXL
  Illustrious XL v0.1 — the base Illustrious anime SDXL model.
• JuggernautXLv8
  Juggernaut XL v8 (RunDiffusion) — earlier, very popular photorealistic SDXL.
• JuggernautXLv9
  Juggernaut XL v9 — top-tier photorealistic SDXL.
• MajicMixRealistic
  majicMIX realistic v7 — photoreal SD1.5 (portraits, Asian aesthetics).
• MeinaMix
  MeinaMix v11 (SD1.5) — extremely popular anime SD1.5.
• NeverEndingDream
  NeverEnding Dream (SD1.5) — anime/semi-real SD1.5.
• NoobAiXL
  NoobAI-XL (NAI-XL) EPS v1.1 — Illustrious-based anime SDXL, uncensored.
• NsfwGenAnime
  UnfilteredAI NSFW-GEN-ANIME — anime-style NSFW SDXL.
• NsfwGenV2
  UnfilteredAI NSFW-gen v2 — SDXL tuned for explicit content, unfiltered.
• NsfwV1
  NSFW v1.0 (SDXL) — the "nsfw-v1" model from LocalAIHub.
• OmnigenXL
  OmnigenXL — NSFW/SFW all-rounder SDXL.
• OmnigenXLNsfw
  OmnigenXL NSFW/SFW v1.0 — the "omnigen-nsfw-v10" model from LocalAIHub.
• PlaygroundV25
  Playground v2.5 — aesthetic-focused SDXL-class model.
• PonyDiffusionV6XL
  Pony Diffusion V6 XL (SPO, diffusers) — hugely popular anime/furry SDXL, uncensored.
• RealVisXL4
  RealVisXL V4.0 — photorealistic SDXL, renders NSFW without filters.
• RealVisXL5
  RealVisXL V5.0 — newest photorealistic RealVis SDXL.
• RealVisXL5Lightning
  RealVisXL V5.0 Lightning — photoreal SDXL, ~6 steps (very fast).
• RealisticVision51
  Realistic Vision v5.1 — classic photoreal SD1.5.
• SdxlBase
  Stable Diffusion XL base 1.0 — the reference SDXL model.
• SdxlTurbo
  SDXL Turbo — single-/few-step SDXL from Stability AI (very fast).
• StableDiffusion15Diffusers
  Stable Diffusion v1.5 (reference SD1.5 checkpoint, diffusers).
• WaiIllustriousSDXL
  waiIllustrious SDXL v15 — anime/illustrious NSFW SDXL.
• value__

KnownHuggingFaceTextModel

enum

Curated chat/LLM models hosted on HuggingFace that are NOT in the LocalAI gallery (the UnfilteredAI uncensored family). Loaded via the generated-config AddTextModel(name, model, backend) — the TEXT counterpart of AddHuggingFaceModel (which is image/diffusers-only). Served on /v1/chat/completions. The DescriptionAttribute carries the load reference: a huggingface://owner/repo/file.gguf URI runs on the robust llama-cpp backend, a bare owner/repo id runs the full safetensors weights on the vllm backend (more VRAM, and vllm must support the architecture — verify on your GPU). All repos verified public/text-generation.

Values

• BadMistral_1_5b
  UnfilteredAI BADMISTRAL-1.5B — tiny uncensored Mistral-class (safetensors → vllm).
• DanL3R1_8b
  UnfilteredAI DAN-L3-R1-8B — uncensored Llama-3 8B ("DAN" reasoning). GGUF f16 → llama-cpp (robust).
• DanQwen35_4b
  UnfilteredAI Dan-Qwen3.5-4B — uncensored Qwen3 4B (safetensors → vllm).
• DanQwen3_1_7b
  UnfilteredAI DAN-Qwen3-1.7B — small uncensored Qwen3 (safetensors → vllm).
• HelveteNano
  UnfilteredAI Helvete-nano — nano uncensored model (safetensors → vllm).
• NsfwFlash
  UnfilteredAI NSFW-flash — uncensored NSFW chat model. GGUF Q4_K_M → llama-cpp (robust).
• UnfilteredAi_1b
  UnfilteredAI UNfilteredAI-1B — 1B uncensored base (safetensors → vllm).
• value__

KnownImageModel

enum

Well-known text-to-image models from the LocalAI model gallery. The DescriptionAttribute holds the exact installable gallery name.

Values

• Chroma1Hd
  Chroma1-HD — 8.9B text-to-image derived from FLUX.1-schnell.
• DreamShaper
  DreamShaper — popular general-purpose SD fine-tune.
• Flux1Dev
  FLUX.1 dev (GGML quantized).
• Flux1DevDiffusers
  FLUX.1 dev (diffusers, full fp16) — reference FLUX.1-dev; large download, heavy VRAM.
• Flux1DevGgmlQ8
  FLUX.1 dev (GGML Q8_0) — higher-quality quant than the default GGML build; more VRAM.
• Flux1DevUncensored
  FLUX.1 dev abliterated v2 (GGML Q8) — uncensored variant without content filtering. Runs on the lightweight stablediffusion-ggml backend (~13 GB download).
• Flux1DevUncensoredDiffusers
  FLUX.1 dev abliterated v2 (full fp16, diffusers backend) — highest quality, but a ~24 GB download and heavy VRAM usage. Prefer Flux1DevUncensored locally.
• Flux1KontextDev
  FLUX.1 Kontext dev — image editing / in-context generation.
• Flux1KreaDev
  FLUX.1 Krea dev (GGML) — photographic aesthetic without the "flux look".
• Flux1KreaDevGgmlQ8
  FLUX.1 Krea dev (GGML Q8_0) — higher-quality quant of the photographic Krea model.
• Flux1Schnell
  FLUX.1 schnell — fast, high-quality (Apache-2.0).
• Flux2Dev
  FLUX.2 dev — newest FLUX generation.
• Flux2Klein4b
  FLUX.2 klein 4B — small and fast FLUX.2 (low VRAM).
• Flux2Klein9b
  FLUX.2 klein 9B — mid-size FLUX.2.
• Ideogram4
  Ideogram 4 (GGML Q8) — very strong text rendering.
• Ideogram4Iq4nl
  Ideogram 4 (GGML IQ4_NL) — smaller/faster Ideogram 4 quant with strong text rendering.
• QwenImage
  Qwen-Image — strong text rendering inside images.
• QwenImageEdit
  Qwen-Image-Edit — instruction-driven image editing (edit an input image from a prompt).
• QwenImageEdit2509
  Qwen-Image-Edit 2509 — updated (Sept 2025) Qwen image-editing model.
• StableDiffusion15
  Classic Stable Diffusion 1.5 (GGML, small and fast).
• StableDiffusion35Large
  Stable Diffusion 3.5 Large (GGML) — high quality, needs more VRAM.
• StableDiffusion35Medium
  Stable Diffusion 3.5 Medium (GGML).
• StableDiffusion3Medium
  Stable Diffusion 3 Medium.
• StableDiffusionAio
  Bundled default of the LocalAI AIO images (no extra download).
• ZImageDiffusers
  Z-Image (diffusers) — full (non-turbo) Z-Image, higher quality than the turbo variant.
• ZImageTurbo
  Z-Image Turbo (diffusers) — very fast generations.
• ZImageTurboGgml
  Z-Image Turbo (GGML) — very fast generations on the ggml backend.
• ZImageTurboVllm
  Z-Image Turbo via the vllm-omni backend — fast generations served through vLLM.
• value__

KnownSoundModel

enum

Well-known text-to-sound / MUSIC generation models from the LocalAI gallery, installed via AddSoundModel. Served on the ElevenLabs-compatible POST /v1/sound-generation (fields model_id + text, plus optional music metadata like lyrics, bpm, duration_seconds). Returns binary audio (wav/flac/mp3). Weights are multi-GB and generation is GPU-bound — combine with WithDataVolume. Any other gallery model with a sound_generation usecase works via the AddSoundModel(string) overload.

Values

• AceStepCppTurbo
  ACE-Step 1.5 Turbo, native C++/GGML build (acestep-cpp backend). Stereo 48kHz; Q8_0 quant for a good speed/quality balance.
• AceStepCppTurbo4b
  ACE-Step 1.5 Turbo (C++/GGML) with the larger 4B LM — higher-quality metadata/code generation.
• AceStepTurbo
  ACE-Step 1.5 Turbo — music generation from text/lyrics with BPM/key/time-signature control (ace-step backend). Good default.
• value__

KnownSpeechToTextModel

enum

Well-known speech-to-text (whisper) models from the LocalAI gallery, installed via AddSpeechToTextModel. Served OpenAI-compatibly on /v1/audio/transcriptions. Names follow the standard whisper.cpp gallery size variants; any other gallery model works via the AddSpeechToTextModel(string) overload.

Values

• Moonshine
  Moonshine (crispASR) — fast, low-latency English speech recognition.
• ParakeetTdt1_1b
  NVIDIA Parakeet TDT 1.1B (crispASR) — fast, accurate English transcription.
• Qwen3Asr0_6b
  Qwen3-ASR 0.6B — small modern multilingual ASR model.
• Qwen3Asr1_7b
  Qwen3-ASR 1.7B — larger Qwen3 ASR, higher accuracy.
• Voxtral
  Voxtral (crispASR) — Mistral's speech-understanding/transcription model.
• WhisperBase
  Whisper base (74M) — good default, fast, low memory.
• WhisperBaseQ5
  Whisper base (Q5_1) — quantized base; very light default.
• WhisperLarge
  Whisper large (v2, 1.55B) — original large multilingual model.
• WhisperLargeQ5
  Whisper large (Q5_0) — quantized large; best-accuracy multilingual at a smaller size.
• WhisperLargeTurbo
  Whisper large turbo — distilled large-v3, much faster with near-large accuracy.
• WhisperLargeTurboQ8
  Whisper large turbo (Q8_0) — quantized turbo; smaller/faster, minimal quality loss.
• WhisperMedium
  Whisper medium (769M) — high accuracy.
• WhisperMediumQ5
  Whisper medium (Q5_0) — quantized medium; good accuracy/size balance.
• WhisperSmall
  Whisper small (244M) — better accuracy, still light.
• WhisperSmallQ5
  Whisper small (Q5_1) — quantized small; light memory footprint.
• WhisperTiny
  Whisper tiny (39M) — fastest, tiniest whisper; lowest accuracy.
• value__

KnownTextModel

enum

Well-known chat / LLM text models from the LocalAI gallery, installed via AddTextModel and served on /v1/chat/completions. Includes vision-capable multimodal (Qwen3-VL, Gemma 3) and an omni model. The gallery holds 1000+ LLMs — this is a curated pick; ANY other works via the AddTextModel(string) overload. MoE sizes like "30b-a3b" = 30B total / 3B active.

Values

• Aya23_35b
  Aya 23 35B — größeres mehrsprachiges Cohere-Modell.
• Aya23_8b
  Aya 23 8B — mehrsprachiges Modell von Cohere.
• Codestral22b
  Codestral 22B v0.1 — starkes Code-Modell.
• DeepHermes3Llama8b
  DeepHermes 3 Llama 3 8B (Preview) — Reasoning-fokussiert.
• DeepHermes3Mistral24b
  DeepHermes 3 Mistral 24B (Preview).
• DeepSeekCoderV2Lite
  DeepSeek-Coder V2 Lite Instruct (MoE) — starkes Coding bei geringem Aktiv-Speicher.
• DeepSeekOcr
  DeepSeek-OCR — Dokument-/Bild-zu-Text.
• DeepSeekR1DistillLlama70b
  DeepSeek-R1 Distill Llama 70B.
• DeepSeekR1DistillLlama8b
  DeepSeek-R1 Distill Llama 8B.
• DeepSeekR1DistillQwen14b
  DeepSeek-R1 Distill Qwen 14B.
• DeepSeekR1DistillQwen32b
  DeepSeek-R1 Distill Qwen 32B.
• DeepSeekR1DistillQwen7b
  DeepSeek-R1 Distill Qwen 7B.
• DeepSeekV3_2
  DeepSeek V3.2 — Top-Reasoning (sehr groß).
• DevstralSmall2507
  Devstral Small 2507 — Coding-/Agent-optimiert.
• Dolphin30Llama31_8b
  Dolphin 3.0 Llama 3.1 8B — hilfsbereiter, unzensierter Allrounder.
• Dolphin30Llama32_3b
  Dolphin 3.0 Llama 3.2 3B — kompakt, unzensiert.
• Dolphin30Mistral24b
  Dolphin 3.0 Mistral 24B (CognitiveComputations).
• Falcon3_10b
  Falcon3 10B Instruct.
• Falcon3_1b
  Falcon3 1B Instruct.
• Falcon3_3b
  Falcon3 3B Instruct.
• Gemma3_12b
  Gemma 3 12B (Vision).
• Gemma3_1b
  Gemma 3 1B.
• Gemma3_27b
  Gemma 3 27B (Vision).
• Gemma3_4b
  Gemma 3 4B (Vision).
• Gemma3n_E2b
  Gemma 3n E2B — kleinere Gemma-3n-Variante.
• Gemma3n_E4b
  Gemma 3n E4B (effiziente MatFormer-Architektur).
• Glm47Flash
  GLM-4.7 Flash — schnelles, starkes Chat-Modell.
• Glm52
  GLM 5.2 — starkes, aktuelles Chat-/Reasoning-Modell.
• GptOss120b
  gpt-oss 120B — großes offenes OpenAI-Modell.
• GptOss20b
  gpt-oss 20B — OpenAIs offenes Modell (mittelgroß).
• Granite32_8b
  IBM Granite 3.2 8B Instruct.
• Granite33_2b
  IBM Granite 3.3 2B Instruct.
• Granite33_8b
  IBM Granite 3.3 8B Instruct.
• Granite3_1b
  IBM Granite 3.0 1B (MoE).
• Granite40HSmall
  IBM Granite 4.0 H Small (Hybrid).
• Granite40HTiny
  IBM Granite 4.0 H Tiny (Hybrid).
• Granite40Micro
  IBM Granite 4.0 Micro — sehr kompakt.
• Hermes3Llama8bLorablated
  Hermes 3 Llama 3.1 8B (lorablated, unzensiert).
• Hermes4_14b
  NousResearch Hermes 4 14B.
• Hermes4_70b
  NousResearch Hermes 4 70B.
• KimiK26
  Kimi K2.6.
• KimiK27Code
  Kimi K2.7 Code — sehr starkes Coding-Modell.
• Llama31_70b
  Llama 3.1 70B Instruct.
• Llama31_8b
  Llama 3.1 8B Instruct.
• Llama32_1b
  Llama 3.2 1B Instruct (Q4).
• Llama32_3b
  Llama 3.2 3B Instruct (Q4).
• Llama33_70b
  Llama 3.3 70B Instruct.
• MagistralSmall2509
  Magistral Small 2509 — Mistrals Reasoning-Modell.
• MiniCpmV45
  MiniCPM-V 4.5 — kompaktes, leistungsstarkes Vision-LLM.
• Mistral7bV03
  Mistral 7B Instruct v0.3 — kompakter, robuster Klassiker.
• MistralNemo2407
  Mistral NeMo Instruct 2407 (12B) — 128k-Kontext, mehrsprachig.
• MistralSmall24b2501
  Mistral Small 24B Instruct 2501.
• MistralSmall31_24b
  Mistral Small 3.1 24B Instruct 2503.
• MistralSmall32_24b
  Mistral Small 3.2 24B Instruct 2506.
• Pixtral12b
  Pixtral 12B — Mistrals multimodales (Vision) Modell.
• Qwen3Coder480b
  Qwen3-Coder 480B-A35B — Spitzen-Coding (sehr groß).
• Qwen3Omni30b
  Qwen3-Omni 30B-A3B Instruct — multimodal (Text/Audio/Bild).
• Qwen3Vl2b
  Qwen3-VL 2B Instruct — kompaktes Vision-LLM.
• Qwen3Vl30b
  Qwen3-VL 30B-A3B Instruct — starkes Vision-LLM (MoE).
• Qwen3Vl32b
  Qwen3-VL 32B Instruct — großes Vision-LLM.
• Qwen3Vl4b
  Qwen3-VL 4B Instruct — Vision-LLM.
• Qwen3Vl8b
  Qwen3-VL 8B Instruct — Vision-LLM.
• Qwen3_0_6b
  Qwen3 0.6B — winzig, für schnelle/lokale Tests.
• Qwen3_14b
  Qwen3 14B.
• Qwen3_235bA22b
  Qwen3 235B-A22B Instruct 2507 (MoE) — sehr großes Flaggschiff.
• Qwen3_30bA3b
  Qwen3 30B-A3B (MoE, 3B aktiv) — stark bei geringem Speed-Kosten.
• Qwen3_32b
  Qwen3 32B.
• Qwen3_4b
  Qwen3 4B.
• Qwen3_8b
  Qwen3 8B — guter Allrounder-Default für eine einzelne GPU.
• SmolLm2_1_7b
  SmolLM2 1.7B — sehr leichtgewichtig.
• value__

KnownTextToSpeechModel

enum

Well-known text-to-speech / audio models from the LocalAI gallery, installed via AddTextToSpeechModel. The DescriptionAttribute holds the exact installable gallery name. Served on /v1/audio/speech (OpenAI-compatible) and /tts. Any other gallery model works via the AddTextToSpeechModel(string) overload.

Values

• Chatterbox
  Chatterbox — expressive TTS with zero-shot voice cloning (Resemble AI).
• Dia
  Dia — 1.6B dialogue TTS (Nari Labs); generates lifelike multi-speaker conversation.
• FishSpeechS2Pro
  Fish-Speech S2 Pro — high-quality multilingual TTS with voice cloning.
• KittenTts
  Kitten-TTS — very small / fast TTS.
• Kokoro
  Kokoro — multilingual (incl. German), fast, high quality. Good default.
• KokoroGerman
  Kokoro (Rust "kokoros"), German voices.
• KokoroMultiLang
  Kokoro multi-language v1.0 (sherpa-onnx) — offline multilingual Kokoro.
• Kokoros
  Kokoros (Rust) — the default multilingual Kokoro voice set.
• KokorosChinese
  Kokoros (Rust) — Chinese (Mandarin) voices.
• KokorosJapanese
  Kokoros (Rust) — Japanese voices.
• Lfm2AudioTts
  LiquidAI LFM2.5-Audio 1.5B — TTS variant of the audio foundation model.
• NeuTtsAir
  NeuTTS Air — lightweight, natural on-device TTS with instant voice cloning.
• OmniVoice
  OmniVoice (cpp) — fast TTS with voice cloning from a reference clip.
• OmniVoiceHq
  OmniVoice (cpp, high quality) — higher-fidelity OmniVoice variant.
• OuteTts
  OuteTTS — multilingual TTS.
• ParlerTtsMini
  Parler-TTS Mini v0.1 — controllable TTS (steer voice/style via a text description).
• PiperEnglishGbAlan
  Piper VITS — English (GB) male voice "Alan".
• PiperEnglishGbVctk
  Piper VITS — English (GB) multi-speaker "VCTK".
• PiperEnglishUsAmy
  Piper VITS — English (US) female voice "Amy".
• PiperFrenchSiwis
  Piper VITS — French voice "Siwis".
• PiperGerman
  Piper — German voice (Thorsten), small & robust offline.
• PiperItalianPaola
  Piper VITS — Italian voice "Paola".
• PiperSpanishDavefx
  Piper VITS — Spanish voice "DaveFX".
• PocketTts
  Pocket-TTS — small, fast general-purpose TTS.
• Qwen3Tts0_6b
  Qwen3-TTS 0.6B — custom-voice TTS (clone a voice from a reference sample).
• Qwen3Tts1_7b
  Qwen3-TTS 1.7B — larger custom-voice TTS, higher quality.
• Supertonic3
  SuperTonic 3 — fast, expressive neural TTS.
• VibeVoice
  VibeVoice — expressive multi-speaker TTS.
• VibeVoiceCpp
  VibeVoice (C++/GGML build) — expressive multi-speaker TTS, native backend.
• VitsLjs
  VITS — English "LJSpeech" (classic LJS single-speaker voice).
• VoiceChineseHuayan
  Piper voice — Chinese (Mandarin) "Huayan", medium quality.
• VoiceEnglishAmy
  Piper voice — English (US) female "Amy", medium quality.
• VoiceEnglishRyan
  Piper voice — English (US) male "Ryan", high quality.
• VoiceFrenchTom
  Piper voice — French male "Tom", medium quality.
• VoiceGermanThorsten
  Piper voice — German (Thorsten), medium quality.
• VoiceGermanThorstenEmotional
  Piper voice — German (Thorsten), emotional variant.
• VoiceItalianPaola
  Piper voice — Italian "Paola", medium quality.
• VoiceRussianDenis
  Piper voice — Russian male "Denis", medium quality.
• VoiceSpanishDavefx
  Piper voice — Spanish "DaveFX", medium quality.
• VoxCpm15
  VoxCPM 1.5 — tokenizer-free, context-aware TTS with voice cloning.
• value__

KnownVideoModel

enum

Well-known text/image-to-video models from the LocalAI gallery, installed via AddVideoModel. Served on POST /video. Video weights are large (many GB) and generation is slow / GPU-bound. The gallery is the source of truth — browse `` and pass any exact id via the AddVideoModel(string) overload.

Values

• Ltx2
  Lightricks LTX-2 — DiT audio-video foundation model; generates synchronized video and audio (image-to-video, diffusers). GPU.
• Ltx23
  Lightricks LTX-2.3 — improved LTX-2 with better motion/quality; synchronized audio-video (diffusers). GPU.
• Wan21FirstLastFrameToVideo720pGgml
  Wan 2.1 first-last-frame-to-video 14B 720p, GGUF Q4_K_M — interpolates between a start and end image (great for seamless loops).
• Wan21ImageToVideo480pGgml
  Wan 2.1 image-to-video 14B 480p, GGUF Q4 — animates a reference image into a short clip.
• Wan21ImageToVideo720pGgml
  Wan 2.1 image-to-video 14B 720p, GGUF Q4_K_M — native 720p single-image animation.
• Wan21TextToVideoGgml
  Wan 2.1 text-to-video 1.3B, GGUF-quantized (stable-diffusion.cpp). Cheapest Wan; ~10 GB RAM, CPU-offloadable.
• Wan22ImageToVideo
  Wan 2.2 image-to-video (14B) via the vllm-omni backend — animates a still image.
• Wan22TextToVideo
  Wan 2.2 text-to-video (14B) via the vllm-omni backend. Needs a strong GPU.
• value__

LocalAiBuilderExtensions

static class

Aspire hosting extension for a self-hosted, OpenAI-compatible multimodal AI service (LocalAI): image generation, text-to-speech, speech-to-text, video generation, chat and embeddings — the self-hosted counterpart of AddOllama for everything beyond text.

LocalAiGpu

enum

GPU vendor for the LocalAI container.

Values

• Amd
  AMD GPU via ROCm devices (/dev/kfd, /dev/dri).
• None
  CPU only (works everywhere, slow).
• Nvidia
  NVIDIA GPU (--gpus all; needs NVIDIA Container Toolkit / Docker Desktop GPU support).
• value__

LocalAiOpenWebUIResource

class

An Open WebUI container wired to a LocalAiResource.

Constructors

• LocalAiOpenWebUIResource(string name)
  An Open WebUI container wired to a LocalAiResource.

LocalAiOptions

class

Options for AddLocalAI.

Constructors

• LocalAiOptions()

Properties

• AioProfile : string { get; set; }
  AIO profile (cpu, gpu-8g, apple). LocalAI's AIO images detect the GPU via lspci, which fails inside Docker Desktop/WSL2 even when --gpus all works — so when Gpu is Nvidia and an AIO tag is used, this defaults to gpu-8g to force GPU mode. Set explicitly to override.
• ApiKey : string { get; set; }
  Optional API key the backend requires (sets LocalAI API_KEY).
• Environment : IDictionary<string, string> { get; }
  Extra environment variables for the container.
• Gpu : LocalAiGpu { get; set; }
  GPU vendor. Default: Nvidia.
• HostPort : int? { get; set; }
  Fixed host port for the endpoint (random if null).
• Image : string { get; set; }
  Container image (without tag). Default: localai/localai.
• Tag : string { get; set; }
  Image tag. Default: the standard NVIDIA CUDA 12 build (currently LocalAI 4.x) — this is what brings video generation and the ace-step sound backend. It's a slim image that only loads what you add via AddModel/AddTextModel/…; backends download on demand. NOTE: the all-in-one (-aio-) tags are frozen at v3.12.1 upstream and lack video/sound — only pick one (e.g. latest-aio-gpu-nvidia-cuda-12) if you specifically want the bundled default model set and can do without the newer backends.

LocalAiResource

class

A self-hosted, OpenAI-compatible multimodal AI service (backend: LocalAI). One container serves image generation (/v1/images/generations), text-to-speech (/v1/audio/speech), speech-to-text (/v1/audio/transcriptions), video generation (/video), sound / music generation (/v1/sound-generation), chat and embeddings — plus LocalAI's built-in WebUI, all on the same endpoint.

Constructors

• LocalAiResource(string name)
  A self-hosted, OpenAI-compatible multimodal AI service (backend: LocalAI). One container serves image generation (/v1/images/generations), text-to-speech (/v1/audio/speech), speech-to-text (/v1/audio/transcriptions), video generation (/video), sound / music generation (/v1/sound-generation), chat and embeddings — plus LocalAI's built-in WebUI, all on the same endpoint.

Methods

• DefaultModelFor(ModelModality modality) : string
  The default model id for a modality (first added of that kind), or null if none was added.

Properties

• DefaultModel : string { get; }
  The default IMAGE model consumers use (first image model added wins; falls back to the AIO-bundled stablediffusion when no image model was added explicitly).
• Models : IList<RegisteredModel> { get; }
  Models registered via AddModel/AddTextToSpeechModel/AddSpeechToTextModel/ AddVideoModel/AddSoundModel (installed from the LocalAI gallery on startup), each tagged with its modality.

Fields

• DefaultTargetPort : int
  Default internal container port LocalAI listens on.
• HttpEndpointName : string
  Name of the primary HTTP endpoint.

ModelModality

enum

The kind of generation a model performs — determines which default-model env var it drives.

Values

• Embedding
  Text embeddings for semantic search / RAG (POST /v1/embeddings).
• Image
  Text-to-image (/v1/images/generations).
• Sound
  Text-to-sound / music generation (POST /v1/sound-generation).
• SpeechToText
  Speech-to-text / transcription (/v1/audio/transcriptions).
• Text
  Chat / LLM text generation — incl. vision-capable multimodal (POST /v1/chat/completions).
• TextToSpeech
  Text-to-speech (/v1/audio/speech).
• Video
  Text/image-to-video (POST /video).
• value__

OpenWebUiOptions

class

Options for the WithOpenWebUI(...) overloads. Defaults reproduce the built-in wiring (LocalAI registered as an OpenAI-compatible connection + image generation, env authoritative). Override to change auth, persistence, the image model/tag, or add arbitrary env.

Constructors

• OpenWebUiOptions()

Properties

• ApiKey : string { get; set; }
  API key sent to LocalAI for the OpenAI/images calls. Default sk-local.
• Auth : bool? { get; set; }
WEBUI_AUTH. null = leave untouched (a reused WebUI keeps its own setting); a newly-created WebUI defaults to false (no login, dev-friendly).
• EnableImageGeneration : bool { get; set; }
  Wire image generation against LocalAI's OpenAI-compatible images endpoint. Default true.
• Environment : IDictionary<string, string> { get; }
  Extra environment variables for the Open WebUI container (applied last — overrides the above).
• Image : string { get; set; }
  Container image — only used when a NEW Open WebUI is created. Default ghcr.io/open-webui/open-webui.
• ImageGenerationModel : string { get; set; }
  Image-generation model Open WebUI uses. Default: the LocalAI default image model.
• PersistentConfig : bool { get; set; }
ENABLE_PERSISTENT_CONFIG. Default false so this env is authoritative on every start — otherwise Open WebUI freezes these values in its DB on first run and ignores later env changes (that's why a reused WebUI wouldn't pick up the config). Set true to let the UI persist changes.
• RegisterOpenAiModels : bool { get; set; }
  Register LocalAI as an OpenAI-compatible model connection (chat/LLM/vision list). Default true.
• Tag : string { get; set; }
  Image tag (new WebUI only). Default main.

RegisteredModel

class

A model queued for install on LocalAI startup, tagged with its ModelModality.

Constructors

• RegisteredModel(string Name, string Reference, ModelModality Modality)
  A model queued for install on LocalAI startup, tagged with its ModelModality.

Properties

• Modality : ModelModality { get; set; }
  The generation kind this model performs.
• Name : string { get; set; }
  The model id consumers pass to the API (and shown in /v1/models).
• Reference : string { get; set; }
  What goes into LocalAI's MODELS list: a gallery name, URI or container path to a generated config yaml.

SdNextResource

class

A standalone SD.Next image-generation studio attached to the stack.

Constructors

• SdNextResource(string name)
  A standalone SD.Next image-generation studio attached to the stack.

Projects

Nextended_Aspire_Hosting_LocalAI

class

Metadata for the Aspire AppHost project.

Properties

• ProjectPath : string { get; }
  The path to the Aspire Host project.

↩ Back to the package page