English · Français · Español · Deutsch · Italiano · Português (Brasil) · Русский · 简体中文 · 日本語 · العربية · हिंदी · বাংলা · اردو · मराठी · తెలుగు · தமிழ் · Kiswahili · Hausa · Tagalog · فارسی (ایران) · አማርኛ · ไทย · Basa Jawa · Bahasa Indonesia · Polski · Tiếng Việt · 繁體中文 · Čeština · 한국어 · Català · ગુજરાતી · ಕನ್ನಡ · Yorùbá · မြန်မာ · Lingála · ଓଡ଼ିଆ · Українська · മലയാളം · ਪੰਜਾਬੀ (ਗੁਰਮੁਖੀ) · Nederlands · Naijíriá Píjin · مصري · پنجابی (شاہ مکھی) · Português (Portugal) · 粵語 · 吴语（上海话） · عربي شامي · عربي سوداني · भोजपुरी · 晋语 · 閩南語 · دارجة جزائرية · 客家話 · الدارجة المغربية · 湘语 · سنڌي · Basa Sunda · Ìgbò · دری · नेपाली · Bahasa Melayu · O‘zbekcha (shimoliy) · isiZulu · عربي صعيدي · پښتو (شمالي) · Oromoo (Giddugaleessa-Lixaa) · سرائیکی · Soomaali · Română · অসমীয়া · Sinugbuanon · Қазақша · සිංහල · isiXhosa · ខ្មែរ · Afrikaans · मैथिली · Wolof · Kurmancî · Ikinyarwanda
You describe. Novera prepares, runs and organizes.
An independent AI assistant by TRON-ABS, from first launch to results: text, images, music and video in one conversation.
Available for Windows, Linux, macOS and Android · 80 interface languages · 42 cloud providers · 49 media presets · Apache 2.0
Automation · Dashboard · Local models · Formats & Hub · Cloud · Français · Contribute
Official contact: contact.novera.tronabs@gmail.com
Official Novera social profiles: Facebook · Instagram · TikTok · YouTube · X · Bluesky
Share your creations, models, settings, questions, feedback and ideas.
Data and privacy: [local inference, optional telemetry, error reports and update connections (French)](docs/DONNEES_ET_CONFIDENTIALITE.md). License: [Apache 2.0](LICENSE) · [Attribution](NOTICE).
---
## Availability — 3 November 2026
**Novera 1.0.0 is available for Windows, Linux, macOS and Android.** [Windows x64 / Linux x64](https://github.com/TRON-ABS/Novera/releases/tag/v1.0.0) · [Windows ARM64](https://github.com/TRON-ABS/Novera/releases/tag/v1.0.0-windows-arm64). Desktop packages, both CLI archives, signatures and update manifests are public and verified.
| Platform | Availability |
| --- | --- |
| Windows x64 | Public NSIS installer, signed updater; native installation and replacement verified. |
| Linux x64 | Public AppImage, signed updater; native installation and replacement verified. |
| Windows x64 / Linux x64 CLI | [Novera CLI](https://github.com/TRON-ABS/Novera/releases/tag/v1.0.0-cli-novera): `novera.exe` / `novera`, with a dedicated signed update channel. |
| Windows ARM64 | [Public NSIS installer](https://github.com/TRON-ABS/Novera/releases/tag/v1.0.0-windows-arm64), native CPU text engine; installation, replacement and public-channel checks verified. |
| Android ARM64 | [Public signed APK 1.0.0](https://github.com/TRON-ABS/Novera/releases/tag/v1.0.0-android-1000012), Android 7+, default batch 512. The maintainer confirms build 1000012 works on their phone. Clean package with no conversations, models or personal API keys. [Evidence and limitations](docs/RELEASE_ANDROID_1.0.0.md). |
| macOS Intel / Apple Silicon | [Official 1.0.0 DMG release](https://github.com/TRON-ABS/Novera/releases/tag/v1.0.0-macos), macOS 13+ Intel / 14+ Apple Silicon; Novera-signed updates. Not Apple notarized; automated checks passed separately on both architectures, visual checks deferred. [Installation and evidence](docs/RELEASE_MACOS_1.0.0.md). |
| iOS | Internal simulator build; Apple signing/provisioning is required before iPhone installation. [Follow-up](docs/REPRISE_MACOS_IOS.md). |
| Windows x64 portable | [Signed ZIP archive](https://github.com/TRON-ABS/Novera/releases/tag/v1.0.0-windows-portable); data stays in Data, signed Program replacement verified. |
| Web, MSI, Debian | Hosted Web client and other formats remain internal; Pages presentation site is available. |
On Mac, open the DMG and copy Novera into Applications. If the first launch is blocked, allow Novera in **System Settings → Privacy & Security → Open Anyway**. [Apple instructions](https://support.apple.com/guide/mac-help/mh40616/mac).
Updater signatures are not Windows Authenticode or Apple notarization. Manual current-version checks pass against the public channel; a transition between two public versions remains a next-release test. [Evidence and limitations](docs/LAUNCH_STATUS.md).
On Android, a modern phone with enough memory is recommended for local models; use an API provider on less powerful devices. For web search, a context of at least **16K tokens (16,384)** is recommended, consistently across versions.
**Windows x64 portable 1.0.0 is available.** Discussions, settings and downloaded models stay inside the portable folder under `Data`. Extract the complete archive and start `Novera.exe`. Close Novera before moving the folder or ejecting USB storage. [Portable guide and limitations](docs/WINDOWS_PORTABLE.md).
## One dashboard for every creation
Novera grows with you. Your examples and settings can help other users find a suitable setup. Share questions, issues and ideas on our [Facebook page](https://www.facebook.com/profile.php?id=61596114515071) or by [email](mailto:contact.novera.tronabs@gmail.com). If possible, include your Novera version, device and local/API model. Remove personal information and never share API keys or passwords.
*Real Windows application screenshot supplied by the maintainer on 1 November 2026, who reports that the application works well. It shows a generated landscape, its saved-file controls and the unified creation dashboard; this is not a mockup.*
One panel, five modes: **Auto / Text / Image / Music / Video**. Controls follow the selected mode: ratio, resolution, duration, cadence, quality, instrumental, lyrics, seed and quantity. Panel width and buttons adapt to the window and translated labels.
- **1–10 sequential generations**, defaulting to 1, with independent quantities for image, music and video. Stop cancels the active task and the remaining queue.
- **Random, fixed, +1 and −1 seeds**, preset or custom ratios and optional **Real-ESRGAN ×2/×4** upscaling with final dimensions displayed.
- The selected model, automatic memory, attachments, instructions, web search and tools stay in the same bar. Tools depend on model capabilities and granted access.
## From first launch to results
**Our automation connects preparation to the complete execution workflow.** Novera handles the technical steps available for the selected configuration, while keeping advanced controls accessible.
| Step | What Novera handles |
| --- | --- |
| **1. Start** | Selects a supported system language at first launch and saves your preference. Presents starting models and useful hardware information. |
| **2. Prepare** | Guided installation offers text, embeddings, image, music, video and upscaling. It previews components, sizes and free space before your approval, then tracks downloads with progress, pause/resume and SHA-256 verification. |
| **3. Configure** | Registers models and their components, prepares the relevant profiles and reports required dependencies. Retrieval embeddings start on CPU with a saved CPU/GPU choice. New music profiles default to instrumental mode. |
| **4. Understand the request** | Auto routes recognized creation requests to text, image, music or video. You can also explicitly select the output mode and a local or cloud model. |
| **5. Execute** | Automatic memory waits for active work, releases unneeded models and reloads required resources. A 1–10 job series runs in order and releases resources before the next job. Cloud providers use the selected account and configuration. |
| **6. Retrieve and reuse** | Saves media in Novera system folders, displays results and available timings, then offers variants, the library, copying, exports or system sharing. Sending to another application remains your action. |
Automation covers integrated engines and workflows; it does not make every Hub model executable. Hardware, installed components and provider access determine what can run. Folders, file changes and shell commands follow the conversation's access grants.
[User guide](docs/USER_GUIDE.md) · [Engines and validation](docs/SUIVI_MOTEURS.md) · [Development](docs/DEVELOPMENT.md)
In the updated sources, **Configurations adapts its recommendations to mobile and constrained hardware**: LFM3-1.7B Q4, or LFM3-Omni-2B Q4 with its projector when reported memory and engine requirements are met. Unavailable local generation is hidden or blocked; configured cloud image, music and video remain available. Interface guidance reuses the 80 language catalogs. These estimates still require device testing and a rebuild to reach installed applications.
## Capabilities in the source
Optional maintenance diagnostics now identify builds and record handled failures as well as JavaScript exceptions, with separate consent and filtered technical fields. An hourly production email alert is configured for the maintainer. Private source maps and alert details follow the [maintenance runbook](docs/MAINTENANCE.md); test email receipt is confirmed, while authenticated symbol upload and testing in the next binary remain release checks.
| Area | Features |
| --- | --- |
| **Text and vision** | Conversation, document attachments, streamed replies and local or cloud models. Vision depends on the model; understanding and generating images are separate capabilities. |
| **Image** | In-conversation generation, seeds, ratios, resolution, GGUF/Safetensors components, configured Diffusers folders and the supported native NVFP4 CUDA profile. |
| **Music** | Local or cloud generation, duration, quality, instrumental or vocals where supported, structured lyrics and an audio player. Provider options differ. |
| **Video** | Local or cloud generation, duration, frame count/cadence for local profiles, ratios, resolution and a video player. |
| **Search and tools** | Web search, embedding-based document retrieval, MCP, workspaces and local tools according to model support and granted access. |
| **Results and sharing** | Media library, variants, response/conversation exports, documents and original media, printing/PDF and platform-dependent system sharing. |
| **Resource management** | Automatic memory enabled by default, CPU/GPU settings for supported components and manual profiles when resources allow. |
## Local models in use
These models form the current guided configuration. You can import other compatible models or change the profiles.
| Role | Model / configured components |
| --- | --- |
| **Text — first-run model** | **Nemotron-3.5-Bolt-E4B**, file `Nemotron-3.5-Bolt-E4B-Q4_K_M.gguf`. |
| **Document retrieval** | **Qwen3-Embedding-0.6B Q8_0**, **CPU by default**, with a saved CPU/GPU choice. |
| **Image** | **FLUX.2 Klein 4B Q4_K_M**, **Qwen3-4B Q4_K_M** encoder and official FLUX.2 VAE. |
| **Image — hardware option** | **FLUX.2 Klein 4B NVFP4**, through the native CUDA path when the GPU and environment are compatible. |
| **Music** | **ACE-Step 1.5 Turbo 2B Q4_K_M**, `acestep-5Hz-lm-1.7B-Q8_0.gguf` planner, dedicated Qwen music encoder and VAE. |
| **Video** | **Wan2.1 T2V 1.3B Q4_K_M**, **UMT5-XXL Q4_K_M** encoder and Wan 2.1 VAE. |
| **Upscaling** | Optional **Real-ESRGAN ×2 and ×4**, with dependency checks. |
The Qwen music encoder and the Qwen retrieval model are **different files**, stored separately. Model licenses remain those of their publishers.
[Text manifest](web-app/src/constants/text-manifest.json) · [Embeddings](web-app/src/constants/embedding-manifest.json) · [Music](web-app/src/constants/music-manifest.json) · [Image/video components](src-tauri/src/core/creation_bundle.rs)
## Cloud providers
**42 provider profiles** are integrated. **ElevenLabs** now appears in the showcase alongside the music routes from Stability AI, fal and Replicate.
| | | |
| --- | --- | --- |
| 
[OpenAI](https://platform.openai.com/docs/models) | 
[Microsoft Azure](https://oai.azure.com/deployments) | 
[Anthropic](https://platform.claude.com/docs/en/models/overview) |
| 
[OpenRouter](https://openrouter.ai/models) | 
[Mistral](https://docs.mistral.ai/models) | 
[Groq](https://console.groq.com/docs/models) |
| 
[xAI](https://docs.x.ai/developers/models) | 
[Google Gemini](https://ai.google.dev/gemini-api/docs/models) | 
[MiniMax](https://platform.minimax.io/docs/api-reference/text-openai-api) |
| 
[Hugging Face](https://huggingface.co/models?pipeline_tag=text-generation&inference_provider=all) | 
[NVIDIA NIM](https://build.nvidia.com/models) | 
[DeepSeek](https://api-docs.deepseek.com/) |
| 
[Kimi · Moonshot AI](https://platform.kimi.ai/docs/models) | 
[Alibaba Cloud · Qwen](https://www.alibabacloud.com/help/en/model-studio/models) | 
[Cerebras](https://inference-docs.cerebras.ai/models/overview) |
| 
[Together AI](https://docs.together.ai/docs/serverless/models) | 
[Fireworks AI](https://docs.fireworks.ai/getting-started/models) | 
[Perplexity](https://docs.perplexity.ai/docs/agent-api/models) |
| 
[Meta AI](https://dev.meta.ai/docs/models) | 
[Cohere](https://docs.cohere.com/docs/models) | 
[Z.ai](https://docs.z.ai/guides/llm/glm-5.3) |
| 
[AI21 Labs](https://docs.ai21.com/docs/jamba-foundation-models) | 
[BytePlus · ByteDance](https://docs.byteplus.com/en/docs/modelark/model-parameter-support) | 
[Baidu Qianfan](https://cloud.baidu.com/doc/qianfan/s/rmh4stp0j) |
| 
[Tencent TokenHub](https://cloud.tencent.com/document/product/1823/130078) | 
[Amazon Bedrock](https://docs.aws.amazon.com/bedrock/latest/userguide/inference-chat-completions.html) | 
[IBM watsonx](https://www.ibm.com/docs/en/watsonx/saas?topic=preview-inferencing-gateway-models) |
| 
[ElevenLabs](https://elevenlabs.io/docs/api-reference/music/compose) | 
[Black Forest Labs](https://docs.bfl.ai/quick_start/generating_images) | 
[Stability AI](https://docs.stability.ai/api-reference) |
| 
[Recraft](https://www.recraft.ai/docs/api-reference/agents.md) | 
[Ideogram](https://developer.ideogram.ai/ideogram-api/api-overview) | 
[Adobe Firefly](https://developer.adobe.com/firefly-services/docs/firefly-api/) |
| 
[Runway](https://docs.dev.runwayml.com/api/) | 
[Luma](https://docs.agents.lumalabs.ai/guides/model/) | 
[LTX (Lightricks)](https://docs.ltx.io/models) |
| 
[fal](https://fal.ai/docs/documentation/model-apis/inference/queue) | 
[Replicate](https://replicate.com/docs/reference/http) | 
[Leonardo.Ai](https://docs.leonardo.ai/reference/creategeneration) |
| 
[Krea](https://www.krea.ai/docs/developers/krea-3/overview) | 
[Vidu](https://platform.vidu.com/docs/model-map) | 
[PixVerse](https://docs.platform.pixverse.ai/changelog-2394996m0) |
### Image, music and video
**20 providers expose generation routes**, with **49 presets: 25 image, 20 video and 4 music**. Music routes are **ElevenMusic v2.5**, **Stable Audio 3**, **ACE-Step through fal** and **Stable Audio 2.5 through Replicate**.
| Provider | Outputs in Novera | Catalog models |
| --- | --- | --- |
| [ElevenLabs](https://elevenlabs.io/docs/api-reference/music/compose) | Music | `ElevenMusic v2.5` |
| [Stability AI](https://docs.stability.ai/api-reference/stable-audio/text-to-audio) | Image / Music | `Stable Audio 3` · `Stable Image Ultra` · `Stable Image Core` |
| [fal](https://fal.ai/models/fal-ai/ace-step/api) | Image / Music / Video | `ACE-Step · cloud` · `Pika 2.2 · fal` · `Qwen Image 2512 · fal` · `FLUX.2 Max · fal` · `Kling 3 Pro · fal` · `Kling 3 Standard · fal` |
| [Replicate](https://replicate.com/stability-ai/stable-audio-2.5/api/schema) | Image / Music | `Stable Audio 2.5 · Replicate` · `FLUX.2 Max · Replicate` |
| [BytePlus · ByteDance](https://docs.byteplus.com/en/docs/ModelArk/image-generation-api) | Image / Video | `Dola Seedream 5 Pro` · `Dreamina Seedance 2.5` |
| [Adobe Firefly](https://developer.adobe.com/firefly-services/docs/firefly-api/api/) | Image / Video | `Adobe Firefly Image 5` · `Adobe Firefly Video` |
| [Google Gemini](https://ai.google.dev/gemini-api/docs/omni) | Image / Video | `Gemini Omni 1.1 Flash` · `Gemini Nano Banana 2.1` |
| [Krea](https://www.krea.ai/docs/developers/krea-3/overview.md) | Image | `Krea 3` |
| [Leonardo.Ai](https://docs.leonardo.ai/docs/lucid-origin) | Image | `Lucid Origin` · `Krea 2 Turbo · Leonardo` · `Ideogram 3 · Leonardo` |
| [Vidu](https://platform.vidu.com/docs/api-reference/video-models/vidu-q4-preview/image-to-video) | Video | `Vidu Q4 Preview` · `Vidu Q3 Pro` · `Vidu Q3 Turbo` |
| [OpenAI](https://developers.openai.com/api/docs/guides/image-generation) | Image | `GPT Image 2.5 Sunburst` · `GPT Image 2.5 Flare` |
| [xAI](https://docs.x.ai/developers/model-capabilities/imagine) | Image / Video | `Grok Imagine Image 2.0` · `Grok Imagine Video 1.5` · `Grok Imagine Video 1.5 Lite` |
| [Black Forest Labs](https://docs.bfl.ai/quick_start/generating_images) | Image | `FLUX.2 [max]` · `FLUX.2 [pro]` · `FLUX.2 [klein] 9B` |
| [Recraft](https://www.recraft.ai/docs/api-reference/agents.md) | Image | `Recraft V4.1` · `Recraft V4.1 Flash` · `Recraft V4.1 Pro` |
| [Ideogram](https://developer.ideogram.ai/ideogram-api/api-overview) | Image | `Ideogram 4.5` |
| [Runway](https://docs.dev.runwayml.com/api/) | Video | `Runway Gen-4.5` |
| [Luma](https://docs.agents.lumalabs.ai/guides/model/) | Image / Video | `Luma Uni-1 Max` · `Luma Uni-1` · `Luma Ray 3.2` |
| [LTX (Lightricks)](https://docs.ltx.io/models) | Video | `LTX-2.5 Pro` · `LTX-2.5 Fast` |
| [MiniMax](https://platform.minimax.io/docs/api-reference/video-generation-v2-create) | Image / Video | `MiniMax H3 Max` · `MiniMax H3` · `MiniMax H3 Highspeed` · `MiniMax Image-01` |
| [PixVerse](https://docs.platform.pixverse.ai/changelog-2394996m0) | Video | `PixVerse V6 Max` · `PixVerse V6` |
Show text and vision model shortlists by provider
| Provider | Model IDs offered in the source |
| --- | --- |
| [OpenAI](https://platform.openai.com/docs/models) | `gpt-6.1-sol` · `gpt-6-astra` · `gpt-6-luna` · `gpt-6-sol` |
| [Microsoft Azure](https://oai.azure.com/deployments) | Account-specific deployment or gateway |
| [Anthropic](https://platform.claude.com/docs/en/models/overview) | `claude-sonnet-5-5` · `claude-opus-5-5` · `claude-fable-5-1` · `claude-haiku-5-5` |
| [OpenRouter](https://openrouter.ai/models) | `moonshotai/kimi-k3.1-pro` · `minimax/minimax-m3.1-flash` · `deepseek/deepseek-v4.1-flash-1017` · `nvidia/nemotron-3.5-super-vl:free` |
| [Mistral](https://docs.mistral.ai/models) | `mistral-large-4` · `mistral-medium-3-5` · `mistral-small-2603` · `ministral-3b-2512` |
| [Groq](https://console.groq.com/docs/models) | `openai/gpt-oss-120b` · `openai/gpt-oss-20b` · `llama-3.3-70b-versatile` · `llama-3.1-8b-instant` |
| [xAI](https://docs.x.ai/developers/models) | `grok-4.7` |
| [Google Gemini](https://ai.google.dev/gemini-api/docs/models) | `gemini-4-argon` · `gemini-3.8-flash` · `gemini-3.6-flash` · `gemini-3.5-flash-lite` |
| [MiniMax](https://platform.minimax.io/docs/api-reference/text-openai-api) | `MiniMax-M3.1-Flash` · `MiniMax-M3` |
| [Hugging Face](https://huggingface.co/models?pipeline_tag=text-generation&inference_provider=all) | `deepseek-ai/DeepSeek-V4.1-Flash-1017` · `Qwen/Qwen3.8-27B` · `moonshotai/Kimi-K3` · `zai-org/GLM-5.3` |
| [NVIDIA NIM](https://build.nvidia.com/models) | `minimaxai/minimax-m3.1-flash` · `deepseek-ai/deepseek-v4.1-flash` · `nvidia/nemotron-3.5-super-vl-120b-a12b` · `nvidia/nemotron-3.5-lightning-30b-a3b` · `z-ai/glm-5.3` |
| [DeepSeek](https://api-docs.deepseek.com/) | `deepseek-flash` · `deepseek-v4-pro` |
| [Kimi · Moonshot AI](https://platform.kimi.ai/docs/models) | `kimi-k3.1-pro` · `kimi-k3.1-pro-highspeed` · `kimi-k3` |
| [Alibaba Cloud · Qwen](https://www.alibabacloud.com/help/en/model-studio/models) | `qwen3.8-max` · `qwen3.7-plus` · `qwen3.8-flash` |
| [Cerebras](https://inference-docs.cerebras.ai/models/overview) | `qwen-3.8-27b` · `gpt-oss-120b` |
| [Together AI](https://docs.together.ai/docs/serverless/models) | `deepseek-ai/DeepSeek-V4.1-Flash-1017` · `moonshotai/Kimi-K3` · `Qwen/Qwen3.8-Flash` · `Qwen/Qwen3.8-2.4T-A95B` |
| [Fireworks AI](https://docs.fireworks.ai/getting-started/models) | `accounts/fireworks/models/deepseek-v4p1-flash-1017` · `accounts/fireworks/models/kimi-k3` · `accounts/fireworks/models/minimax-m3` |
| [Perplexity](https://docs.perplexity.ai/docs/agent-api/models) | `perplexity/sonar` · `openai/gpt-6.1-sol` · `google/gemini-4-argon` · `perplexity/kimi-k3` |
| [Meta AI](https://dev.meta.ai/docs/models) | `muse-wonder-1.0` · `muse-spark-1.3` |
| [Cohere](https://docs.cohere.com/docs/models) | `command-a-plus-05-2026` |
| [Z.ai](https://docs.z.ai/guides/llm/glm-5.3) | `glm-5.3` · `glm-5.3-flash` · `glm-5.3-flashx` |
| [AI21 Labs](https://docs.ai21.com/docs/jamba-foundation-models) | `jamba-large` · `jamba-mini-2` |
| [BytePlus · ByteDance](https://docs.byteplus.com/en/docs/modelark/model-parameter-support) | `dola-seed-2-1-turbo-260628` |
| [Baidu Qianfan](https://cloud.baidu.com/doc/qianfan/s/rmh4stp0j) | `ernie-5.1` |
| [Tencent TokenHub](https://cloud.tencent.com/document/product/1823/130078) | `hy4-preview` |
| [Amazon Bedrock](https://docs.aws.amazon.com/bedrock/latest/userguide/inference-chat-completions.html) | `openai.gpt-6.1-sol` |
| [IBM watsonx](https://www.ibm.com/docs/en/watsonx/saas?topic=preview-inferencing-gateway-models) | Account-specific deployment or gateway |
Inventory checked against the source on **29 October 2026**. These names describe Novera presets, not guaranteed access for every account. Remote catalogs, regions, licenses, quotas and billing remain provider-specific. Compatible profiles can refresh their catalogs; Azure and some gateways use account deployments. Cloud contracts have simulated tests; no paid generation was performed for this showcase update.
[Media catalog](web-app/src/constants/mediaProviders.ts) · [Text/vision models](web-app/src/constants/cloudModels.ts) · [Cloud setup](docs/USER_GUIDE.md#cloud) · [Icon provenance](docs/DEVELOPMENT.md#logos)
## Hub and compatible models
**GGUF, standard and quantized Safetensors, ONNX, and MLX**: explore models by task and choose variants compatible with the engine and your hardware.
| Family | Support in the source | Requirements |
| --- | --- | --- |
| **GGUF** | Text through llama.cpp; image/video components for supported profiles. | Architecture recognized by the bundled engine. |
| **Split GGUF** | One variant per complete series, summed size, all-shard download and size/SHA-256/header checks. | Standard numbered filenames; downloading 500 GB does not mean the model fits this PC. |
| **Standard Safetensors** | Complete folders through Transformers, causal text on CPU or NVIDIA CUDA. | Configuration, tokenizer and every required shard. |
| **ONNX Runtime** | Local import of numeric/bool graphs and prepared GenAI text exports. | Validated complete folder; task-specific preprocessing/postprocessing required; [setup and limits](onnx-server/README.md). |
| **LoRA / PEFT adapters** | Import a complete base with its separate local adapter in the Safetensors provider. | Unquantized causal Transformers base, matching identity, tensor checks and cumulative memory estimate; [setup and limits](transformers-server/README.md#local-lora--peft-adapters). |
| **FP8 LLM / ModelOpt** | Native CUDA execution for compatible dense Llama/Qwen2/Qwen3 per-tensor FP8 exports. | Architecture/tensor checks and real kernel probe; [setup and limits](transformers-server/README.md#modelopt-fp8-and-nvfp4-llms). |
| **NVFP4 LLM / ModelOpt** | Packed block-16 FP4 projections through native CUDA GEMMs; separate from image NVFP4. | Compatible Blackwell GPU, supported dimensions/export and actual kernel probe; untrained test models tested on RTX 5070. |
| **AWQ** | Unfused 4-bit GEMM checkpoints. | NVIDIA CUDA and optional GPTQModel/Transformers loaders. |
| **GPTQ** | 2/3/4/8-bit configurations, gptq/gptq_v2 formats. | NVIDIA CUDA, GPTQModel and Optimum; real execution tested at 4 bits. |
| **bitsandbytes** | 4-bit NF4/FP4 or 8-bit INT8. | NVIDIA CUDA and optional dependencies; real NF4 and INT8 tests. |
| **EXL3** | Complete quantized Safetensors folders through ExLlamaV3; real CUDA generation tested. | NVIDIA GPU, separate Python 3.12 environment and compatible architecture; no CPU fallback. |
| **EXL2** | Dedicated ExLlamaV2 loader; streaming, cancellation and unload tested on GPU. | Explicitly select the CUDA Attention backend for the tested software stack; a complete model checkpoint and sufficient available VRAM are required. |
| **MLX** | Compatible models on Apple Silicon. | MLX engine and supported platform. |
The Hub now pins **LFM3-1.7B** as its LiquidAI recommendation for compatible configurations, with Q4_K_M, Q6_K and Q8_0 variants. The first-run model remains the one listed in the guided configuration above.
Filter download weights from **under 1 GB to 500 GB and above**, then sort by hardware fit, size, recency or downloads. Results load progressively; filters apply to fetched models and their default variant. Memory advice uses available RAM/VRAM, context and companion components, without double-counting shared GPU memory.
AWQ 4-bit, GPTQ 4-bit, NF4 and INT8 each generated text on an RTX 5070 Laptop using tiny untrained models created offline. This validates loading and execution, not the quality or performance of every model. **These additions are in the source: rebuilding and compatible dependencies are required; this README update does not upgrade an installed binary.**
**Measured on the RTX 5070 Laptop:** Qwen2.5-0.5B-Instruct, context 2,048, prompt 272 tokens, 128 generated tokens, five warmed runs: median decode **GGUF Q4_K_M 402.7 tokens/s · EXL3 4 bpw 312.5 · EXL2 4 bpw 146.9**. No speed benefit over GGUF was demonstrated here. Quantization/calibration differs, and EXL3 failed a basic arithmetic check; these results are not a general speed or quality ranking. [Full measurements and limits](deliverables/exllama-2026-10-30/validation.md).
**TabbyAPI** is an optional, separately managed OpenAI-compatible server connection, disabled by default. Authenticated EXL3 model discovery and streaming were tested. The managed local ExLlama route remains distinct. [ExLlama setup and TabbyAPI connection](transformers-server/README.md#exl3-first-exl2-second--nvidia-execution).
The default hardware-adapted sort favors newer releases within the same compatibility and available-memory tier. Installed models are not replaced automatically.
[Quantized Safetensors setup](transformers-server/README.md#quantized-safetensors--source-integration-29-october-2026) · [Formats, priorities and validation](docs/SUIVI_MOTEURS.md)
The updated sources add **optional managed components in Configurations for Windows x64**: private Python environments, pinned CPU/CUDA profiles, estimates and confirmation before downloads, verification, repair and removal. Custom interpreter paths stay under advanced options. Native GGUF engines keep their existing components; mobile and cloud do not download desktop Python/PyTorch. Three private CPU/export profiles were tested using synthetic test data; the maintainer reports successful desktop use on 1 November 2026; GPU profiles and clean-PC installation still need separate, detailed validation. [Profiles and test limits](docs/REPRISE_COMPOSANTS_GERES_80_LANGUES.md). Bundling these components in the EXE is deferred.
## 80 interface languages and variants
**All 80 catalogs cover the current 2,429 English keys** across 15 namespaces, including menus, configurations, settings, help and media controls. Plural forms and interface direction follow the selected language. Response and voice languages depend separately on the model.
This technical coverage does not certify linguistic quality: **native review and regional adaptation remain necessary**. See native labels and known limits in the [language guide](docs/LANGUAGES.md) and [audit report](deliverables/interface-audit-2026-10-29/RAPPORT.md).
## Helping people, in every sector
Beyond school and education, our ambition is to help people learn, solve problems, create and develop useful projects worldwide. Novera aims to make current AI advances easier to understand and use, across languages, levels of experience and available hardware.
| Sectors we want to help develop | Examples of useful contributions |
| --- | --- |
| Schools, education and lifelong learning | Explain concepts, adapt learning materials and support teachers, students and families. |
| Research, science and space | Explore literature, explain methods, analyse data and share knowledge. |
| Health, care and social support | Make information easier to understand and support research and professional workflows. |
| Agriculture, food and rural communities | Share practical knowledge, support planning and develop locally useful tools. |
| Environment, climate, water and biodiversity | Explain evidence, support observation and help communities plan projects. |
| Energy, industry, construction and crafts | Support documentation, design, maintenance, technical learning and process improvement. |
| Businesses, commerce, finance and employment | Help organise work, understand information, develop projects and learn professional skills. |
| Public services, law and civic participation | Make documents and procedures easier to understand and information easier to access. |
| Computing, cybersecurity, robotics and telecommunications | Help develop software, understand systems and build useful tools. |
| Arts, culture, media, languages and heritage | Support writing, translation, images, video, creative projects and cultural sharing. |
| Transport, logistics, tourism, sport and community life | Help with planning, communication, learning and local initiatives. |
| Accessibility, humanitarian work and everyday life | Reduce language and knowledge barriers and support people with practical assistance. |
These are directions for contribution and development, not a claim that every sector-specific application is already delivered or validated. We want simple guidance for everyday use, clear advanced options when needed, transparent hardware and cloud requirements, and collaboration with the people who know each field and local context.
## Project status
**Novera 1.0.0 is available for Windows, Linux, macOS and Android.** The screenshot shows a Windows application compiled locally by the maintainer. Source changes, targeted tests and hardware validation are documented separately; this does not validate every model, platform or cloud account.
Manual GitHub Actions builds are enabled for all eleven distribution variants. Follow the [build results and verified candidate downloads](docs/CI_DISTRIBUTIONS.md); availability is recorded separately for each platform. Candidates are distinct from signed production releases. Distribution requirements are described in the [official release policy](OFFICIAL_RELEASES.md) and [releases page](https://github.com/TRON-ABS/Novera/releases). Existing updater-disabled candidates remain internal. Public releases require a configured and verified update mechanism; [desktop signature checks and remaining launch requirements are documented](docs/LAUNCH_STATUS.md).
## Documentation and contributions
- [Detailed user guide](docs/USER_GUIDE.md), [product and user-information golden rules (French)](docs/REGLES_D_OR.md), [models and engines](docs/SUIVI_MOTEURS.md).
- [Development](docs/DEVELOPMENT.md), [build and release guide](docs/BUILD_AND_RELEASE.md), [future versions and update preparation](docs/FUTURES_VERSIONS.md).
- [Project dashboard](TABLEAU_DE_BORD_NOVERA.md), [roadmap](ROADMAP.md), [changelog](docs/CHANGELOG.md).
- [Contributing](CONTRIBUTING.md): code, tests, native review across 80 languages and real workflow feedback.
## Identity and license
Novera is maintained by **TRON-ABS** in [this repository](https://github.com/TRON-ABS/Novera). See the [identity policy](BRANDING.md) and [official release policy](OFFICIAL_RELEASES.md).
Novera builds on [Jan](https://github.com/janhq/jan), created by Menlo Research and its contributors. Original and third-party attributions and the [Apache 2.0 license](LICENSE) are retained. Jan downloads are not Novera releases.
## Support Novera
As an independent researcher, I know how a lack of funding can make it difficult to pursue research.
If you would like to be part of the Novera adventure, you can support my research and the project's development with a voluntary bitcoin contribution.
Thank you to everyone for your support, contributions, feedback and encouragement. Every gesture counts in this adventure!
**Bitcoin (BTC) — Bitcoin network**
Receiving address: `17TXpB1aBj5DxjwEfM3K44zyqhixRzY78i`
To contribute, scan this QR code with your wallet or copy the address above. Select **Bitcoin (BTC), on the Bitcoin network**, then choose any amount you wish.
---
To everyone beginning their journey with Novera: may your curiosity open new doors, your ideas come to life and every creation take you a little further.
TRON-ABS
Mr Sachet Jean-Marc — Independent researcher — France
The art of Human with AI
Thank you to everyone who contributes to a better world. Through their discoveries, work, mutual support and the knowledge they pass on, they have helped us reach where we are today.
We wish them the very best for the future, along with everyone who carries this spirit of sharing and helping others forward.
Development tool: 🧠 Codex — GPT-6.1 Sol