The hub started as a list of chat models. It's now 56 models from 26 labs, and the biggest growth is in three places: 3D, code and speech. Here's what's there, how to pick, and what we're adding next.
3D: six ways to turn a photo into a mesh
Every 3D model runs on your own GPU and writes a .glb you can drop into Blender, Godot, Unity or a browser.
- TripoSR (MIT): the quickest, around 6 GB. Good for a first try.
- Hunyuan3D-2 mini, mini fast and turbo: Tencent's smaller variants, about 5–6 GB, trading a little detail for speed.
- Hunyuan3D-2: the full shape model, about 6 GB.
- Hunyuan3D-2.1: the newest and most detailed, about 10 GB.
arcflare gen 3d chair.png -m hunyuan3d-2miniIn the desktop app you can then brush over part of the mesh and type what to change, like "make the legs longer", and it edits the model.
Code: from a laptop to a server rack
There are 16 coding models now, and the range is the point. Pick by your GPU:
| Your GPU | Try | Why |
|---|---|---|
| 6–8 GB | Mellum 4B, Seed-Coder 8B, Qwen2.5-Coder 7B | fast completion and small fixes |
| 12–16 GB | DeepSeek-Coder-V2-Lite, Devstral Small 2 | real agent work on one card |
| 20–24 GB | Qwen3-Coder 30B-A3B, GLM-4.7-Flash | strong agents; only ~3B active per token, so they're quick |
| 24 GB+ | OlympicCoder 32B, NextCoder 32B | competition-style problems, careful edits |
| 40 GB+ | Kimi-Dev 72B, Qwen3-Coder-Next 80B-A3B | the big ones, on a Mac with lots of memory or two cards |
Not sure what fits? Models spots your GPU, recommends the best coder for it under the Code tab, and hides everything too big.
Speech: five voices, all local
Qwen3-TTS, Kokoro, Chatterbox, VoxCPM2 and OuteTTS. Kokoro is tiny (82M, about 1 GB) and quick enough for anything; Qwen3-TTS and Chatterbox can clone a voice from a short clip.
arcflare gen tts "Hello from my own GPU" -m kokoroWhat's next
We're adding models as they're released and as we can test them properly: each one gets checked for what it needs, how it's licensed and whether it actually runs before it goes in the hub. On the list:
- more image-to-3D options, including ones that texture the mesh
- more small coding models that are useful on 8 GB cards
- more speech models, and more languages
- image generation models you can run straight from
arcflare shop
Want a specific model? Ask for it: pick Idea and name the model. That list gets read before every update.
arcflare shop and the desktop app's Models view always show the current hub, so new models appear there without updating anything.