Nativ

The #1 local AI platform built on Apple Silicon. Try Nativ:

[Nativ @Nativ_AI]

Part of what makes Nativ awesome is the feedback we get from the community.

If you’ve got any thoughts, bugs, or ideas, drop them below. We’re always listening!

👇

[Nativ @Nativ_AI]

@aarifkhan4u 🚀

[Nativ @Nativ_AI]

Flint and Phonon now run in Nativ. Local coding + on-device voice, all on your Mac 📷

Nativ v0.4.3:
→ Recursive Flint: 40B MoE, day 0 on Apple Silicon
→ Gradium Phonon TTS: 4 voices, first audio under 15 ms
→ New Phonon deep dive on the blog
→ Smoother model picker
→ Snappier chat sidebar
→ Free local AI. Nothing leaves your Mac

Try Nativ: https://t.co/aoAv7sx9nQ

[Nativ @Nativ_AI]

@entretokens Coming on the next release on Monday 🫡

[Nativ @Nativ_AI]

@luismolinaab @Prince_Canuma @Recursive_SI Yep, you’ve got it right, those two are all there is. The 40 in 40B is the parameter count, not GB. And rec does need an M4 Pro or up, but Flint itself doesn’t: we’ve run the 4-bit on an M3 Ultra with mlx-vlm 0.7.10.

[Nativ @Nativ_AI]

@Spangler3000 @Recursive_SI Can't wait to see them! For reference, our M3 Ultra does about 90 tok/s on the 4-bit, and around 420 tok/s total with 8 requests going at once. Curious how far the M5 Ultra pushes that 👀

[Nativ @Nativ_AI]

Congrats to @Recursive_SI on Recursive Flint! Day 0 support in Nativ 📷

40B MoE · 2.5B active · 256K context · fully on-device.

On an M3 Ultra in full BF16:
⚡️ 2,641 tok/s prefill · 51 tok/s decode
🚀 Batch 8 lifts decode to 236 tok/s
📏 Still 33 tok/s decode at 128K
🔓Open weights under Apache 2.0

A 40B coding model that never sends your code anywhere.

Download for macOS 📷
https://t.co/3TQ0Kg4zEE

[Nativ @Nativ_AI]

@AiSimonThompson @RichardSocher @ollama @lmstudio @UnslothAI @exolabs @sgl_project @vllm_project @opencode @pidotdev Here's one. Flint has native MLX support in mlx-vlm 0.7.10:

pip install -U mlx-vlm

The model code and its README: https://t.co/nipUqhkFor

It loads the original BF16 weights as they are.

[Nativ @Nativ_AI]

@lvrpiz @RichardSocher @ollama @lmstudio @UnslothAI @exolabs @sgl_project @vllm_project @opencode @pidotdev Yes! Flint has day-0 MLX support through mlx-vlm 🚀

[Nativ @Nativ_AI]

@jaxcore @GradiumAI Glad it's working now, and thanks for the numbers 🙏 10× on MLX vs up to 22× on WebGPU on an M1 Max is a gap worth looking at. We've passed it to the team. Which mlx-audio version are you on, and what text length?

[Nativ @Nativ_AI]

@JaynitMakwana @RichardSocher @Recursive_SI In our runs Flint peaks at about 23 GB in 4-bit and 81 GB in full BF16 (M3 Ultra, mlx-vlm), and a 128K prompt adds about 4 GB. So for 4-bit we'd pick a 48 GB Mac.

[Nativ @Nativ_AI]

@RichardSocher @ollama @lmstudio @UnslothAI @exolabs @sgl_project @vllm_project @opencode @pidotdev Thanks for having us on day 0 🙏

Flint runs in Nativ, free on Mac. Launch OpenCode or Pi on it from the Integrations page, all local.

In 4-bit it needs about 23 GB and decodes 91 tok/s, or 419 tok/s across 8 requests at once (M3 Ultra, mlx-vlm 0.7.10).

https://t.co/inlCna32fZ

[Nativ @Nativ_AI]

4-bit Flint runs in under a third of the BF16 memory: 23.1 GB at batch 1, 24.7 GB at batch 8. Decode goes from 90.6 tok/s on one request to 419.2 tok/s across 8. M3 Ultra, mlx-vlm 0.7.10, 1K tokens per request.

[Nativ @Nativ_AI]

Long context holds up. Decode goes from 51.3 tok/s at 1K to 32.7 tok/s at 128K, and prefill peaks at 2,641 tok/s at 4K. Single request, M3 Ultra, BF16.

[Nativ @Nativ_AI]

Run several coding agents at once: batch 8 lifts aggregate decode to 235.5 tok/s at 1K, and it still gives 71.3 tok/s at 128K. Same Mac, same BF16 weights.

[Nativ @Nativ_AI]

@voicearena_ai @GradiumAI Thanks for including MLX-Audio 🙏 On a Mac: pip install -U mlx-audio, then python -m mlx_audio.tts.generate --model gradium/phonon-31m-en --text "Hello" --play. Add --stream to hear it as it generates. For the 40M models, accept their terms on Hugging Face first.

[Nativ @Nativ_AI]

@jaxcore @GradiumAI It's already there 🙌 Phonon runs in mlx-audio 0.6.1. On the Mac: mlx_audio.server --host 0.0.0.0 --port 8000. Then any device on your network can POST to /v1/audio/speech (OpenAI-compatible) with model gradium/phonon-31m-en. Add "stream": true to hear it as it's made.

[Nativ @Nativ_AI]

❤️

[Wolfbane @zzddfge]

Loving everyday @Nativ_AI

[Nativ @Nativ_AI]

Congrats to @GradiumAI on Phonon! 🎉 Proud to be partners.

Here it is fully on-device on iPhone 17 Pro and Mac, powered by mlx-audio. Sound on 🔊

40M-parameter voices in English and French. Qwen translates in 1.1s, Phonon speaks both, and nothing leaves the device.

[Gradium @GradiumAI]

A new frontier for on-device speech.

We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date.

Phonon-31m is available under Apache 2.0.

https://t.co/d3qa9qmnzY🧵

[Nativ @Nativ_AI]

Peak memory under 265 MiB: 236 MiB at 8 words, 264 MiB at 1,024.

About a quarter of a gigabyte, light enough to leave running next to everything else on your Mac.

[Nativ @Nativ_AI]

Congrats to @GradiumAI on Phonon! Day 0 support in Nativ 🎉

English text-to-speech · local AI · fully on-device.

On an M5 Max with Nativ v0.4.1, one voice stream:

-> first audio in 10.8 to 14.7 ms
-> 12.5× to 14.1× faster than real time
-> peak memory under 265 MiB
-> benchmarked from 8 to 1,024 words

First audio in under 15 ms, and nothing leaves your Mac.

Download for macOS💻

https://t.co/3TQ0Kg4zEE

Github repo 🛠

https://t.co/tg3OzuBoST

[Nativ @Nativ_AI]

First audio in under 15 ms, from 8 words to 1,024.

Phonon also runs 12.5× to 14.1× faster than real time, so a minute of speech takes under 5 seconds to make.