Krafton releases Raon-1 mini preview, a 30B-A4B MoE model with 262,144 token context
KRAFTON AI announced a preview of its Raon-1 mini language model, a 30B-A4B Mixture‑of‑Experts system built from scratch. The model uses 18 linear attention layers and 7 full attention layers, and it supports up to 262,144 tokens of context. It runs on a single NVIDIA RTX 5090 GPU (32 GB VRAM) and offers quantization in MXFP8 and NVFP4 formats.
Key points
- Raon-1 mini is a 30B-A4B MoE model with 18 linear and 7 full attention layers
- Supports up to 262,144 tokens of context and runs on a single NVIDIA RTX 5090 (32 GB VRAM)
- Sparse MoE activates 6 of 128 experts plus 1 shared, using 4.3B of the 28.8B parameters per token
Raon-1 mini’s sparse MoE activates six of 128 routed experts plus one shared expert per token, using 4.3B of the 28.8B total parameters. The preview includes a DFlash2 draft model for speculative decoding, delivering up to 2.18× higher output throughput on a single NVIDIA B300 compared with autoregressive decoding. The model is released under the Apache 2.0 license, with open‑weight weights, and can be accessed via a free limited‑time playground or through hosted API access after agreeing to separate Hosted Access Terms. The release is labeled a preview, and the article notes that specs and results may change before the final version.
KRAFTON/Raon-1-mini-preview · Hugging Face
huggingface.co · 21 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Nvidia launches Nemotron 3.5 Super VL and Bolt models for enterprise agents and AI PCs · 2 src
- Google and Meta release frontier models to challenge OpenAI and Anthropic · 2 src
- Three benchmarks name different top AI models: Gemini 4 Argon, Claude Opus 5.5, GPT-6.1 Sol · 1 src
- Google unveils Gemini 4 Argon with million-token output window · 3 src
- Google makes Gemini 4 Argon generally available to AI Pro and Ultra subscribers · 38 src
Comments
via GitHub Discussions