# Neywa Labs

> Source: https://aiwiki.ai/wiki/neywa_labs
> Updated: 2026-10-28
> Fact-checked: 2026-10-28
> Categories: AI Companies, AI Inference, Developer Tools, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Neywa Labs." aiwiki.ai, 28 Oct 2026. https://aiwiki.ai/wiki/neywa_labs
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Neywa Labs** is a startup that builds software for running [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) models locally on personal devices, with a focus on Macs and other [Apple silicon](https://aiwiki.ai/wiki/apple_silicon) hardware. It was announced on 27 October 2026, when co-founder and CEO Prince Canuma wrote on X that the company was "coming out of stealth" and described its work as "infrastructure to build and deploy AI that runs on the devices people already own."[1] The company's products are built on mlx-vlm and mlx-audio, two open-source Python libraries that Canuma created on top of Apple's [MLX](https://aiwiki.ai/wiki/mlx) framework, and on Nativ, an open-source macOS app for chatting with and serving local models.[2][5][7][10]

## Overview

Neywa's website describes the company as building "the infrastructure that makes frontier open models run natively on personal devices", with work spanning language, vision, speech, video, model optimization, inference and native applications.[2] It splits that work into open-source engines for developers and products for end users, and says open-source development is "central to how Neywa works."[2] The homepage pitches the stack as a way to run open models and AI agents "privately on Apple Silicon" without a cloud call or API key, and claims the company is "trusted by more than 7M developers."[3] A Devoxx Belgium 2026 speaker biography for Canuma gives a smaller figure for the same libraries, "over 4M combined downloads" for mlx-vlm, mlx-audio and mlx-embeddings.[4]

The company also offers model labs a porting service. According to its site, this can cover architecture implementation and model bring-up, [quantization](https://aiwiki.ai/wiki/quantization) and memory optimization, multimodal inference, MLX support, performance validation across devices, and "day-one support for new model launches."[3]

The name comes, according to the company page, "from Muaneywa", described as "a thing you hang out to dry that never dries" and "a symbol of tenacity."[2]

## History

Canuma published the first version of mlx-vlm in April 2024: the GitHub repository was created on 16 April 2024 and version 0.0.1 reached PyPI on 23 April 2024.[5][6] The mlx-audio repository followed on 27 November 2024, with its first PyPI release on 28 February 2025.[7][8]

The company's GitHub organization, now named Neywa-Labs, was created on 24 August 2025 (the earlier Marvis-Labs address of its main repository now redirects there) and hosts Marvis TTS, a streaming text-to-speech model whose model card lists Prince Canuma and Lucas Newman as its creators and gives a release date of 26 August 2025.[11][12] Nativ, the company's desktop app, appeared on GitHub under Canuma's personal account on 20 July 2026, when version 0.0.1 was published; the repository had 26 releases by 28 October 2026, the latest being v0.4.4.[9]

The company's X account was created on 26 October 2026, and Canuma announced Neywa Labs the next day.[1][22] Canuma's launch post gave the company's mission as to "make AI available to everyone in the world",[1] and a follow-up in the same thread argued for local AI on the grounds that "The model runs on your own device, and your data never leaves it."[23] The Devoxx Belgium 2026 program lists Canuma, under the Neywa Labs affiliation, for a keynote titled "Half of a billion Idle GPUs: The Case for On-Device AI" and a 50-minute talk on running frontier open models on consumer hardware.[4]

## Founders

The company page names three co-founders.[2]

| Name | Role | Background |
|---|---|---|
| Prince Canuma | Co-founder, CEO | Creator of mlx-vlm and mlx-audio.[5][7] His Devoxx biography says he previously led developer experience at Neptune AI and worked at [Arcee AI](https://aiwiki.ai/wiki/arcee_ai) on post-training and inference optimization.[4] His X profile lists Kraków, Poland as his location.[24] |
| Sebastien Michaud | Co-founder, COO | Described by the company as having a career "shaped by commercializing unreleased solutions to critical problems."[2] |
| Lucas Newman | Co-founder, CTO and head of research | Co-creator of Marvis TTS.[12] As of 29 October 2026 he was the top contributor by commit count to mlx-audio (417 commits) and the second-largest to mlx-vlm (306 commits).[25][26] |

## Products and open-source projects

| Project | Type | License | First public release | GitHub stars (29 Oct 2026) |
|---|---|---|---|---|
| mlx-vlm | Python library for vision-language and omni models | MIT | April 2024 | about 5,700[5] |
| mlx-audio | Python library for speech and audio models | MIT | February 2025 (PyPI) | about 8,100[7] |
| Marvis TTS | Text-to-speech model | Apache 2.0 | 26 August 2025 | about 380[11][12] |
| Nativ | macOS desktop app | MIT | 20 July 2026 | about 1,560[9][10] |

### mlx-vlm

mlx-vlm is a package for inference and fine-tuning of [vision language models](https://aiwiki.ai/wiki/vision_language_model) and "Omni Models (VLMs with audio and video support)" on Macs using MLX.[5][6] Its README documents a command-line interface, a FastAPI server with continuous batching and automatic prefix caching, KV-cache quantization, speculative decoding, distributed inference and fine-tuning.[5] Version 0.7.11 was released on PyPI on 28 October 2026.[6]

Several other projects depend on it. [LM Studio](https://aiwiki.ai/wiki/lmstudio)'s open-source MLX engine lists mlx-vlm, alongside mlx-lm and Outlines, as one of the libraries it is built with, for vision model inference,[19] and Docling, the open-source document-conversion toolkit led by IBM Research, includes mlx-vlm as an optional dependency for its inline vision-language models on Apple silicon Macs.[16]

### mlx-audio

mlx-audio is a library for text-to-speech, [speech recognition](https://aiwiki.ai/wiki/speech_recognition) and speech-to-speech models built on MLX.[7][8] Its README also lists music generation, a web interface and API server, quantization, and a Swift package.[7]

### Marvis TTS

Marvis TTS is a conversational speech model for real-time voice cloning and streaming speech synthesis. Its model card says it can clone a voice from 10 seconds of reference audio and that the quantized model is about 500 MB.[12] The first version uses a 250M-parameter multimodal backbone and a 60M-parameter audio decoder, built on the CSM-1B architecture from [Sesame](https://aiwiki.ai/wiki/sesame) and Kyutai's Mimi codec.[12] The card puts the total training cost at "~$2,000", including about $1,500 of experimentation, with pretraining on one NVIDIA GH200.[12] Checkpoints are published under the Marvis-AI organization on [Hugging Face](https://aiwiki.ai/wiki/hugging_face).[12]

### Nativ

Nativ is a native SwiftUI macOS app that bundles an mlx-vlm server and finds compatible models in the user's Hugging Face cache.[10] It requires Apple silicon and macOS 26 or later.[10] According to its README, it provides local chat with image input, a model library with memory-fit warnings, image generation, performance analytics, a system monitor, and OpenAI-compatible and Anthropic Messages-compatible local API endpoints.[10] It can also configure terminal coding agents such as [Claude Code](https://aiwiki.ai/wiki/claude_code), [OpenAI Codex](https://aiwiki.ai/wiki/openai_codex), Aider and Goose to use models it serves.[10]

Neywa's product page says Nativ recommends partner models from Google, [Cohere](https://aiwiki.ai/wiki/cohere) and [Liquid AI](https://aiwiki.ai/wiki/liquid_ai) suited to the user's hardware, and that the app has "no accounts to create, no credits to buy."[13] The site lists a "Nativ Pro" tier as "coming soon."[3] DevoxxGenie, a coding-assistant plugin for JetBrains IDEs, added Nativ as a dedicated local provider in version 1.10.1; a July 2026 post by its creator, Stephan Janssen, described the app's OpenAI- and Anthropic-compatible server on port 8080 by default.[20]

The company's capabilities page also advertises video understanding, voice dictation, a voice mode, meeting recording and transcription, document and image understanding, a personal knowledge base, tools that act on files and the terminal, clustering several devices to pool memory, and fine-tuning.[14]

### Engine and model conversions

Neywa markets its libraries together as an "omni engine" for applications that "read, see, hear, speak, and create directly on personal hardware."[3] Converted model weights are published in a Hugging Face organization named nativ-community, which held 152 models on 29 October 2026.[21]

## Model ports and benchmarks

The company's blog publishes porting write-ups with its own measurements. These figures are self-reported by Neywa Labs.

| Date | Subject | Reported result | Hardware |
|---|---|---|---|
| 26 Oct 2026 | IBM Research's Docling with Granite Docling 258M through mlx-vlm | One page in 1.73 s versus 15.92 s with PyTorch 2.12.1; a four-page batch in 2.45 s versus 22.45 s (about 9.2x faster in both cases)[15] | MacBook with M3 Max |
| 28 Oct 2026 | [OpenBMB](https://aiwiki.ai/wiki/openbmb)'s MiniCPM-OCR 1.0 (a 1.3B document parser) and Jev (a 2B classifier that checks OCR output against the source, published as MiniCPM-OCR-Jev) ported to mlx-vlm | An invoice parsed in 0.381 s; checkpoints published in BF16 and seven quantized formats[17] | M5 Max, 128 GiB |
| 28 Oct 2026 | NVIDIA's Nemotron-3.5-Transcribe-30B-A4B-EA running in Nativ through mlx-audio | A 53.8-second clip processed in 1.59 s with 4-bit weights (33.8x real time), peak MLX allocation 20.37 GB[18] | M3 Ultra, 512 GB |

In the Docling post, Neywa says IBM Research and Neywa Labs "continue to work together" so that new Docling models run on Apple silicon through mlx-vlm when they are released.[15] The MiniCPM-OCR post cautions that its 16x compression mode changed the model's output, so the speed difference is "not an accuracy-equivalent speedup",[17] and the [Nemotron](https://aiwiki.ai/wiki/nemotron) post notes that some quantized runs "added a short utterance absent from the input."[18]

## See also

- [MLX](https://aiwiki.ai/wiki/mlx)
- [On-device AI](https://aiwiki.ai/wiki/on_device_ai)
- [Edge AI](https://aiwiki.ai/wiki/edge_ai)
- [LM Studio](https://aiwiki.ai/wiki/lmstudio)
- [Ollama](https://aiwiki.ai/wiki/ollama)
- [llama.cpp](https://aiwiki.ai/wiki/llama_cpp)
- [Open-source AI](https://aiwiki.ai/wiki/open_source_ai)

## References

1. Prince Canuma, "Today we're coming out of stealth, introducing Neywa Labs", X, 27 October 2026. https://x.com/Prince_Canuma/status/2115128700379566155
2. Neywa Labs, "About Neywa Labs: Making Frontier AI Native" (company page). Retrieved 29 October 2026. https://www.neywalabs.ai/company
3. Neywa Labs, "Local AI native to your Apple device" (homepage). Retrieved 29 October 2026. https://www.neywalabs.ai/
4. Devoxx Belgium 2026, "Prince Canuma" (speaker page). Retrieved 29 October 2026. https://m.devoxx.com/events/dvbe26/speaker/16387/prince-canuma
5. Blaizzy/mlx-vlm, GitHub repository and README. Retrieved 29 October 2026. https://github.com/Blaizzy/mlx-vlm
6. mlx-vlm, Python Package Index (release history). Retrieved 29 October 2026. https://pypi.org/project/mlx-vlm/
7. Blaizzy/mlx-audio, GitHub repository and README. Retrieved 29 October 2026. https://github.com/Blaizzy/mlx-audio
8. mlx-audio, Python Package Index (release history). Retrieved 29 October 2026. https://pypi.org/project/mlx-audio/
9. Blaizzy/nativ, GitHub releases. Retrieved 29 October 2026. https://github.com/Blaizzy/nativ/releases
10. Blaizzy/nativ, GitHub repository and README. Retrieved 29 October 2026. https://github.com/Blaizzy/nativ
11. Neywa-Labs, GitHub organization. Retrieved 29 October 2026. https://github.com/Neywa-Labs
12. Neywa-Labs/marvis-tts, GitHub repository and README (model card). Retrieved 29 October 2026. https://github.com/Neywa-Labs/marvis-tts
13. Neywa Labs, "Nativ" (product page). Retrieved 29 October 2026. https://www.neywalabs.ai/products/nativ
14. Neywa Labs, "Capabilities". Retrieved 29 October 2026. https://www.neywalabs.ai/capabilities
15. Neywa Labs, "Docling x MLX-VLM: how IBM Research runs document AI 9.2x faster on a Mac", 26 October 2026. https://www.neywalabs.ai/blog/docling-mlx-vlm-case-study
16. docling-project/docling, pyproject.toml (models-vlm-inline extra). Retrieved 29 October 2026. https://github.com/docling-project/docling/blob/main/pyproject.toml
17. Neywa Labs, "MiniCPM-OCR on MLX-VLM: document parsing on Apple silicon", 28 October 2026. https://www.neywalabs.ai/blog/minicpm-ocr-mlx-vlm
18. Neywa Labs, "Nemotron Transcribe on Nativ: speech understanding on Apple silicon", 28 October 2026. https://www.neywalabs.ai/blog/nemotron-transcribe-nativ
19. lmstudio-ai/mlx-engine, GitHub README. Retrieved 29 October 2026. https://github.com/lmstudio-ai/mlx-engine
20. Stephan Janssen, "Nativ: A Native macOS Home for Your Local MLX Models, Now a First-Class DevoxxGenie Provider", DevoxxGenie blog, 29 July 2026. https://genie.devoxx.com/blog/nativ-local-mlx-apple-silicon
21. nativ-community, Hugging Face organization. Retrieved 29 October 2026. https://huggingface.co/nativ-community
22. Neywa Labs (@Neywalabs), X profile. Retrieved 29 October 2026. https://x.com/Neywalabs
23. Prince Canuma, "Today, a single AI agent can cost tens of thousands a month to run", X, 27 October 2026. https://x.com/Prince_Canuma/status/2115128704149995967
24. Prince Canuma (@Prince_Canuma), X profile. Retrieved 29 October 2026. https://x.com/Prince_Canuma
25. Blaizzy/mlx-audio, GitHub contributors. Retrieved 29 October 2026. https://github.com/Blaizzy/mlx-audio/graphs/contributors
26. Blaizzy/mlx-vlm, GitHub contributors. Retrieved 29 October 2026. https://github.com/Blaizzy/mlx-vlm/graphs/contributors