ye

Building something that matters;
Foundation Model Research @ KraftonAI

[GooGZ AI @PaulGugAI]

KRAFTON Raon-1 Mini (30B A4B) Preview: a seriously capable LLM for a little agent.

@Kangwook_Lee asked how it performs, so here it is on my RTX 3080 10GB. IQ4_XS. Running the forked llama.cpp build Codex set up for me.

Speed: 1,417 tok/s prompt processing, 45 tok/s generation. Snappy.

The light agent tests were near-perfect:

A1 light agent: 10/10
T1 tool use: 30/30
C1 16K context search: 8.7/10

PA Core 48K (mean/100):
Bonsai 2 27B ternary: 84.3
Raon-1 Mini Preview 30B A4B: 77.6
Qwen 3.6 35B A3B: 60.3

PA2 (scheduling and drafting) was the standout at 99.3, among the best I've seen.

The weak spot was PA4, change of plan (57.3). It spotted the new conflict and the right fix, but it didn't reliably turn that into a complete, consistent handover.

For a mini preview, that's a strong start. Nice work @Krafton_AI👏

[Kangwook Lee @Kangwook_Lee]

Many of you asked us to update the comparison chart with newer models and agentic tasks, so here it is!

[KRAFTON AI @Krafton_AI]

Raon-1 mini is now available for download on Hugging Face without an access request. Thanks for the feedback!

[ye @yvvlim]

Raon-1 mini preview runs super fast while staying competitive with or beating larger korean models

[ye @yvvlim]

i was responsible for the agentic coding behavior of the model, and raon 1 mini is pretty powerful with issue patching / bug fixing tasks! stay tuned!

[Benz ON ⚙️ @Benz_onttv]

@Kangwook_Lee Ok, we may have finaly the real heir to Qwen3.6:35b

Impressive RAW performance. 100% at SWE bench samples. With descent speed at high context, Q6 qant.

[Byungwoo Jeon @rootyjeon]

Can explicitly learning future point motion improve the model that generates those future images?

We introduce VPM (Video-Point Model), a 4D world model that jointly forecasts future multi-view videos + persistent 3D point trajectories.

Short answer: yes. Teaching a video world model to forecast 3D point tracks makes its videos more faithful & 3D-consistent, and makes robots better.

🌐 Project: https://t.co/srVtzZsY5c
🧵 [1/6]

[ye @yvvlim]

@Minseok96_kr Gracias 😇😇

[ye @yvvlim]

hey guys me and my team worked so hard to finally release Raon-1 mini preview
it’s just the beginning, and we are pushing our boundaries to reach frontier intelligence

[Kangwook Lee @Kangwook_Lee]

👋 KRAFTON's Raon is entering the open-weight scene!

Meet Raon-1 mini, our first model, trained from scratch.

On par with leading models in its size class, while running faster and using less memory.

Available under the Apache 2.0 license.
https://t.co/fOFoUcfDLc

[KRAFTON AI @Krafton_AI]

Today, we’re releasing a preview of Raon-1 mini, a 30B MoE LLM with 4B active parameters, trained from scratch by KRAFTON AI.

With our efficient architecture, it matches Muse Glimmer on coding/agentic tasks while being much faster and more memory-efficient! 🔥 Plus, natural, human-like conversation.

This is just a first look. The final release (and bigger models) are coming soon! 👀

👉 Try it now, free for a limited time: https://t.co/uSrLFIGPWy
🤗 Open weights + DFlash drafter (Apache 2.0): https://t.co/XWVmxsDRei