Skip to content
Eugeny

Every model, native on your Mac.

Engines written for Apple silicon. Play the samples right here.

Install./install.sh <model>~/Models/<model>-current

Flagships.

Eugeny

  • 28B params
  • 8.7 GB
  • 262K context

Powerful all-purpose Mac intelligence with vision, latent thinking and calibrated decision heads.

  • Vision
  • Decision heads
  • Speculative drafter
  • Media conditioning

Read the model

A green bookshop front with a sign reading Open until 9 in the left window.

AskIs this shop still open at eight tonight?

Yes, open until nine.

Eugeny warm (M5 Max)
173 ms
GPT-6.1 Sol
3.06 s

Scintilla

  • 3.7 GB INT8
  • 0.54s ready
  • 131K context

The everyday multimodal model, sped up by coupled-tree speculative decoding.

  • Multimodal
  • 58 ms KV restore
  • Media conditioning

AskMove Friday’s call with Giulia to Monday morning.

Done. I drafted a note to Giulia proposing Monday at 10.

Fornace Fast Image

  • 0.5s on M5
  • 2-step DiT
  • Apple silicon

A MOSS fine-tune rewritten for Apple silicon that turns text into images in half a second.

  • 0.5s generation
  • Metal 4.1
  • Eugeny and Scintilla heads
A glossy red ceramic mug with rising steam on a sunlit oak table.
0.5s · M5

H3 Native

  • 90s on Mac
  • 24 fps
  • Audio and video

Generate synchronized 24 fps video and sound directly on your Mac in 90 seconds.

  • Character references
  • Spoken dialogue
  • First and last frame

Fornace Voice

  • Up to 14x realtime
  • 44.1 kHz
  • 3.25 GB

Blazing-fast Mac voice with an Italian dubbing fine-tune, lifelike emotions and natural delivery.

  • Italian fine-tune
  • Breaths and laughter
  • Multi-speaker dialogue
  1. Elena[gasps] Pensavo ci volessero almeno dieci minuti, e invece ha già finito tutto quanto, [laughs] è una cosa davvero incredibile!
  2. Marco[sighs] Non ci credevo neanch'io stamattina, ma appena l'abbiamo provato sul tavolo, [laughs] è partito subito a tutta velocità!
0.00s

Speech, restoration and transcription.

Re-Use Native

  • 40 MB
  • 0.3s per 10s
  • 44.1 kHz

Voice restoration that strips noise and reverb while keeping laughs and breaths.

0.00s

Orukeet Native

  • 864 MB
  • 25 languages
  • Word timestamps

Realtime Metal transcription and 8-speaker diarization: 4 minutes of speech in 1.2 seconds.

Ho prenotato la barca. Partiamo domani mattina alle sei.

0.00s

Fornace Pocket

  • 1.4 GB
  • 24 kHz duplex

Full-duplex realtime conversation with 27 voice presets and instant cloning.

  • Duplex speech
  • 27 presets
  • Instant cloning

EARS Native

  • 389 MB
  • 3.8 ms per 5s

Scores voice quality, 40 emotions and timed laughs, breaths and sighs.

  • Quality
  • 40 emotions
  • Vocal events

Kim-Vocals Native

  • 914 MB
  • 10 min in 6.4s

Studio Mel-Band RoFormer vocal and dialogue isolation in C++20 and Metal 4.1.

  • Vocals
  • Dialogue
  • Stems

ThinkSound and AuK

ThinkSound Native

  • 5.8 GB
  • Seed-exact SFX

Text to foley and ambience, up to 30s on Metal 4.

AuK 1.5B

  • 8.3 GB
  • 24 kHz

Voice design from a description, cloning and speech editing.

Vision, super-resolution and matting.

Qwen-Image 2.1 Native

  • 9.3 GB INT8
  • 10 references
  • True RGBA

Studio image generation and editing with ten ordered references and real transparency.

An open antique brass compass rendered with a transparent background.

VOSR 2.0 Native

  • 1.8 GB W8A8
  • 1-step DiT
  • Up to 16K

One-step super-resolution and detail restoration up to 16,384 pixels.

Low resolution source: a blurry bicycle basket against a yellow wall.

WeVisOne-VL-2B with Mask2Alpha

  • 4.9 GB + 309 MB
  • 0.08 to 0.6s
  • Resident socket

Plain-English grounding, OCR boxes and hair-accurate alpha mattes in one pass.

Source photo: a long haired orange tabby cat on a grey sofa.