Raon-1 mini (KRAFTON) Requirements

Raon-1 mini preview needs roughly 16-20GB for weights at 4-bit and about 18-23GB of memory in practice, so it fits a 24GB GPU such as the RTX 3090 or 4090. It is a 28.8B-total, ~4.3B-active MoE. Using the full 262,144-token context adds KV cache and leaves little headroom on 24GB, but short-to-medium contexts fit (memory figures are estimates from file sizes and the vendor's KV-cache numbers).

Raon-1 mini preview is an Apache 2.0 open-weight model that KRAFTON's AI team released on Hugging Face on October 21, 2026. It was trained from scratch and uses KRAFTON's own linear attention. As the name says, it is a preview: the model card warns that specs and results may change before the final version, which is planned to ship with a technical report (as of October 2026).

If you want to check quickly whether it runs on your GPU or Mac, our free VRAM calculator (no signup) estimates required VRAM from the model, quantization and context length. Raon-1 mini is available in its model list.

Except for BF16, weight sizes are the actual Hugging Face file totals; the memory column adds KV cache and runtime headroom, so it is an estimate. KRAFTON states the 7 full-attention layers use about 8KiB of KV cache per token (BF16), which is only about 2GiB at 262,144 tokens (estimate). The linear-attention layers keep a fixed-size state, so memory grows slowly with context. There is no official GGUF below 4-bit; one community Q2K_Q4 exists, but we could not verify its size or quality. For choosing a quantization, see our GGUF quantization guide.

Because it is an MoE, memory is set by the 28.8B total parameters, not the 4.3B active. Active parameters only reduce per-token compute (speed); unused experts still have to sit in VRAM or RAM. We explain the same idea in our Gemma 4 requirements guide.

All numbers are KRAFTON's own, judged by gpt-5.6-luna, and independent verification is still thin. The model card also compares Qwen3.8-27B (dense) and Muse Glimmer 30B (dense); Raon's MMLU Pro is the lowest of the five (76.5 versus 82.9-85.9 for the others). It also trails the compared MoE models on Terminal-Bench Lite and BFCL v4, so it does not win across the board. Chatbot Korean is an unreleased internal benchmark. The model card asks for an access request (auto-approved), although the API showed the repo as ungated when we checked.

There is little evidence either way. KRAFTON's announcement focuses on Korean and English, and while the model card lists Japanese as a supported language, it publishes no Japanese benchmark. Given the lower knowledge score (MMLU Pro), test it on your own data before relying on it for Japanese factual Q&A. Korean and English work, and agentic use with tool calling, have better published backing. For Japanese-first work, compare it side by side with models that have a proven track record at this size.

As of October 2026, both vLLM and llama.cpp require KRAFTON's forks, since upstream support is pending. For llama.cpp, build the fork and use the GGUF repo.

The official example uses --ctx-size 262144, but on a 24GB GPU it is safer to start smaller, as above. For vLLM, use the NVFP4 checkpoint; this is the officially documented setup for an RTX 5090 (32GB). It needs CUDA Toolkit 13 (nvcc), and the first start spends about 10 minutes compiling kernels.

- Ollama: the model card lists ollama run hf.co/KRAFTON/Raon-1-mini-preview-gguf:Q4_K_M, but since upstream llama.cpp does not support the model yet, we could not confirm it works on stock Ollama. If it fails, use the forked llama.cpp
- LM Studio / Jan: the card advertises support but gives no settings; unverified
- Apple Silicon (MLX): use -mlx (NVFP4, 17.6GB); the tested versions are mlx 0.32.4, mlx-lm 0.32.2 and transformers 5.20.1
- Transformers: pinned to transformers 5.16.1 and accelerate 1.15.0, with trust_remote_code=True; for a 32GB GPU the card suggests the MXFP8 model with CPU offload
- Recommended sampling: temperature 1.0, top_p 0.95, top_k 20; presence_penalty 1.5 for chat and agents, 0 for coding and math; do not disable thinking
- Speculative decoding: DFlash2 (-dflash2) reports up to ~2.18x output throughput on a B300 with vLLM (hardware dependent)

Raon-1 mini has the largest active-parameter count here (~4.3B), needs less memory than Flint, and a bit more than Gemma 4 26B-A4B. By its own numbers it leads on GPQA, IFBench and SWE-Verified, but it is weaker on knowledge and tool calling. Since all five are Apache 2.0, the choice comes down to use case (Korean, agents, Japanese factual Q&A) rather than license.

- Stock llama.cpp or Ollama will not load it: the custom architecture (raon1_mini) is not upstream yet. Build KRAFTON's fork (branch raon-1-mini-preview)
- vLLM's first start is slow: it compiles kernels for about 10 minutes; wait for the logs rather than assuming it hung
- Out of memory: lower the context from 262,144 to something like 32,768, use an fp8 KV cache, or switch to a smaller quantization (IQ4_XS)
- Verbose or repetitive output: check presence_penalty; Transformers has no built-in presence penalty, so the model's remote code applies it
- Reasoning mixed into or missing from output: vLLM returns it in message.reasoning and llama.cpp in message.reasoning_content
- Download refused: the model card says to request access and run hf auth login (approval is automatic)
- It is a preview, so results can change: weights and specs may differ in the final release; wait for it or design for easy swapping