Raon-1-mini-preview Reasoning-Enabled Language Model: 24GB+ VRAM

Raon-1-mini-preview Reasoning-Enabled Language Model: 24GB+ VRAM

At a Glance

Item Value
Repository KRAFTON/Raon-1-mini-preview
Family guide Raon-1 guide (1 articles)
Published 2026-10-21
License apache-2.0
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

KRAFTON AI has released “Raon-1-mini-preview", a Mixture-of-Experts (MoE) language model designed and trained from scratch. It features a total parameter count of 28.8B with 4.3B active parameters per token, and is reported to excel in Korean conversational capabilities and agent tasks. Positioned as a preview version rather than a final release, the model comes with a built-in thinking mode by default.

Specifications

The specifications listed in the model card are as follows:

  • Total parameters: 28.8B
  • Active parameters: 4.3B
  • Architecture: Hybrid MoE (18 Raon Delta Attention layers and 7 Gated MLA layers)
  • Number of layers: 25
  • Number of experts: 128 (6 routed experts + 1 shared expert activated per token)
  • Expert intermediate size: 1,088
  • Hidden Size: 2,560
  • Attention heads: 40 (common to RDA and Gated MLA)
  • Attention head dimensions: 128 (RDA), 192 QK / 128 V (Gated MLA)
  • Query / KV LoRA rank: 576 / 512 (Gated MLA)
  • Positional encoding: NoPE (no positional encoding)
  • Activation function: SiTU-GLU
  • Residual connection: Raon-mHC (4 streams)
  • Vocabulary size: 259,840
  • Context length: 262,144 tokens

Performance

The model card publishes benchmark comparison results against same-scale MoE models (Qwen3.6-35B-A3B, Gemma 4 26B-A4B) and Dense models. In all evaluations, gpt-5.6-luna was used as the judge and user simulator.

→ Scroll horizontally to see all columns

Category Benchmark MoE / Raon-1 mini preview (30B-A4B) MoE / Qwen3.6-35B-A3B MoE / Gemma 4 26B-A4B Dense / Qwen3.8-27B Dense / Muse Glimmer 30B
Knowledge MMLU Pro 76.5 85.9 82.9 85.7 84.0
Knowledge AA-Omniscience 3.5 −7.2 −46.6 −0.5 −34.7
Reasoning GPQA Diamond 84.3 83.2 80.7 89.7 82.6
Reasoning AIME 2026 93.3 92.9 87.1 98.8 92.9
Reasoning HMMT Feb 2026 84.1 84.5 79.9 95.8 74.6
Instruction and multi-turn IFBench 84.2 66.6 80.9 82.4 78.7
Instruction and multi-turn MultiChallenge 59.0 58.2 58.4 63.6 63.4
Coding LiveCodeBench v6 81.6 81.2 83.4 79.3 77.4
Coding SciCode 38.7 35.6 39.6 44.6 43.6
Coding SWE-Verified¹ 70.8 63.2 61.6 74.6 72.0
Coding Terminal-Bench Lite 47.0 63.3 52.0 67.0 70.7
Agentic GDPVal² 66.0 67.4 57.5 63.5 67.3
Agentic AutomationBench³ 29.7 24.9 36.2 58.7 32.1
Agentic BrowseComp Plus 55.7 34.8 32.2 58.0 52.8
Agentic BFCL v4 Overall 59.8 64.7 65.8 72.4 51.8
Long context AA-LCR 67.3 69.3 67.0 80.3 81.3
Long context GDP.pdf⁴ 5.8 6.6 3.6 11.2 8.4
Korean conversation Chatbot Korean (internal)⁵ 92.9 74.5 90.6 90.1 87.0

Additionally, throughput measurement results using “DFlash2", a speculative decoding draft model that accelerates inference, have been published (NVIDIA B300 single, vLLM environment).

→ Scroll horizontally to see all columns

Benchmark AL AR · C=1 DFlash2 · C=1 AR · C=8 DFlash2 · C=8 AR · C=32 DFlash2 · C=32
GSM8K 4.76 296.3 646.3 (2.18×) 1,110.4 1,836.3 (1.65×) 1,528.9 2,619.3 (1.71×)
MATH-500 4.43 308.0 666.6 (2.16×) 1,496.1 2,511.3 (1.68×) 3,082.5 5,166.9 (1.68×)
HumanEval 3.96 305.4 607.9 (1.99×) 1,470.8 2,269.6 (1.54×) 2,793.1 4,596.4 (1.65×)
MBPP 3.85 307.0 604.1 (1.97×) 1,450.6 2,197.2 (1.51×) 2,956.2 5,017.9 (1.70×)
MT-Bench 3.24 304.9 570.2 (1.87×) 1,477.1 2,038.2 (1.38×) 2,521.3 3,935.4 (1.56×)

Based on the publisher’s measurement results, the following points can be gathered regarding the model’s strengths and weaknesses.

First, in the reasoning and mathematics domains, it recorded 93.3 on AIME 2026, which measures US Math Olympiad qualifiers, and 84.3 on GPQA Diamond, which handles PhD-level difficult scientific questions, achieving the highest figures among the comparative MoE model group. It also showed good results with 84.2 on IFBench, which measures instruction-following, and 81.6 on LiveCodeBench v6 using new competitive programming tasks, while also recording a high score of 92.9 on the internal Chatbot Korean benchmark.

On the other hand, it remained at 76.5 on MMLU Pro, which measures general knowledge and reasoning with higher difficulty across 10 choices, falling below other comparative models (82.9 to 85.9). Furthermore, it scored 47.0 on Terminal-Bench Lite, which measures the ability to execute commands on a terminal and complete tasks, and 59.8 on BFCL v4 Overall, which measures the accuracy of function selection and argument construction, falling short of other MoE baselines in some agent and tool-use tasks.

Regarding speculative decoding, it is reported that using DFlash2 in combination improves throughput by up to 2.18x on GSM8K, which solves elementary school-level arithmetic problems, and up to 2.16x on MATH-500, which handles high school-to-competition level mathematics.

Strengths and Use Cases

Raon-1-mini-preview excels at logical reasoning assuming a thinking mode, autonomous agent workflows, and multilingual conversation centered around Korean.

According to the model card, the model has been trained in Korean, English, Japanese, Chinese, Spanish, French, and Thai, featuring fluent conversational ability and practical English performance particularly in Korean. Because it natively supports a massive context length of 262,144 tokens, it is also suited for long-text reading comprehension referencing huge documents or tasks requiring the retention of long-term conversation history.

Additionally, tool calling is built-in as a feature for agent tasks, and integration with agent execution environments such as Hermes Agent, Pi, and OpenCode is officially guided. It demonstrates solid performance in instruction-following, search, and coding fields, making it suitable for complex task processing that combines deep reasoning via thought processes, programming, and the automated execution of various tools.

As recommended generation sampling parameters, it is suggested to enable the thinking mode and set temperature to 1.0, top_p to 0.95, and top_k to 20. It is also specified to set presence_penalty to 1.5 for chat, summarization, knowledge QA, and agent tasks, while setting presence_penalty to 0.0 for coding and math tasks.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 28.8B parameters

Your VRAM Quantization File size Est. memory needed
24GB (RTX 4090 / 3090, etc.) Q4_K_M 18.3GB 22.0GB
32GB (RTX 5090, etc.) Q6_K 24.1GB 28.9GB
48GB (RTX 6000 Ada / A6000, etc.) Q8_0 28.6GB 34.3GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: not registered, vLLM: not registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build KRAFTON/Raon-1-mini-preview-gguf. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but KRAFTON/Raon-1-mini-preview-gguf provides a GGUF build you can use.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compression: the Q4_K_M build measures 5.46 bits per weight — about 34% the size of the original 16-bit weights, calculated by this site from the actual file sizes.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
LiquidAI/LFM2.5-24B-A2B-Base 24.0B 80GB lfm1.0 LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-FP8 24.0B 32GB lfm1.0 LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-MLX-bf16 24.0B 80GB lfm1.0 LFM2.5-24B-A2B-MLX-bf16 Reasoning Model for Apple Silicon: 80GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-MLX-8bit 24.0B 32GB lfm1.0 LFM2.5-24B-A2B-MLX-8bit Text Generation Model: 32GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B 24.0B 12GB lfm1.0 LFM2.5-24B-A2B Agent-Optimized MoE Reasoning Model: 12GB+ VRAM (2026-11-17)

How to Get It

Weights for Raon-1-mini-preview are released under the Apache 2.0 license. The base model is distributed in safetensors format (MXFP8), alongside a variety of formats such as NVFP4 format, GGUF format for llama.cpp, and MLX format for Apple Silicon. Note that the Hugging Face repository is gated, and you must request access on the model page (requests are approved automatically). Authentication using huggingface-cli or hf auth login must be completed before downloading.

Usage methods for major inference environments are as follows:

Running with Transformers

To load and run directly from Hugging Face Transformers in a Python environment, install the dependency libraries as follows, allow remote code execution, and load the model.

pip install "transformers==5.16.1" "accelerate==1.15.0"
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "KRAFTON/Raon-1-mini-preview"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "Tell me about KRAFTON AI."}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    enable_thinking=True,
    truncate_history_thinking=False,
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=32768,
    do_sample=True,
    temperature=1.0,
    top_p=0.95,
    top_k=20,
    min_p=0.0,
    repetition_penalty=1.0,
    presence_penalty=1.5,
)

Since the thinking process is output separated by </think> tags within the generated text, the structure is designed to split and extract strings before and after the tags.

Serving with vLLM

When operating as an OpenAI-compatible API server, use the custom fork of vLLM provided by the publisher. It can be started with tool calling and speculative decoding via DFlash2 enabled.

VLLM_REF="raon-1-mini-preview"
pip install "vllm @ git+https://github.com/krafton-ai/vllm.git@${VLLM_REF}"

vllm serve KRAFTON/Raon-1-mini-preview \
    --host 127.0.0.1 \
    --port 8000 \
    --served-model-name Raon-1-mini-preview \
    --max-model-len 262144 \
    --max-num-batched-tokens 16384 \
    --max-num-seqs 128 \
    --gpu-memory-utilization 0.85 \
    --kv-cache-dtype fp8_e4m3 \
    --enable-prefix-caching \
    --spec-method dflash \
    --spec-model KRAFTON/Raon-1-mini-preview-dflash2 \
    --spec-tokens 7 \
    --reasoning-parser raon_1 \
    --enable-auto-tool-choice \
    --tool-call-parser raon_1 \
    --default-chat-template-kwargs '{
      "enable_thinking": true,
      "truncate_history_thinking": false
    }'

A profile targeting NVFP4 format checkpoints (KRAFTON/Raon-1-mini-preview-nvfp4) is also provided.

Running with llama.cpp

When running the GGUF version using llama.cpp, build from the corresponding dedicated branch. Use KRAFTON/Raon-1-mini-preview-gguf for the checkpoint.

git clone -b raon-1-mini-preview https://github.com/krafton-ai/llama.cpp
cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release -j

After building, it can be run as a CLI chat tool or an OpenAI-compatible server. Model files are downloaded automatically by specifying them as arguments.

./build/bin/llama-cli \
    -hf KRAFTON/Raon-1-mini-preview-gguf:Q4_K_M \
    --ctx-size 262144 \
    --n-predict 32768 \
    --temp 1.0 \
    --top-p 0.95 \
    --top-k 20 \
    --min-p 0.0 \
    --repeat-penalty 1.0 \
    --frequency-penalty 0.0 \
    --presence-penalty 1.5 \
    --reasoning on \
    --chat-template-kwargs '{
      "enable_thinking": true,
      "truncate_history_thinking": false
    }'

Running with MLX

On Apple Silicon Macs, by utilizing the specially converted KRAFTON/Raon-1-mini-preview-mlx, it is possible to launch an OpenAI-compatible endpoint via Python scripts or mlx_lm server.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-11-17 KRAFTON GGUF (imatrix) KRAFTON/Raon-1-mini-preview-gguf Q4_K_M 22.0GB (fits in 24GB VRAM)
2026-11-17 KRAFTON MLX KRAFTON/Raon-1-mini-preview-mlx MLX 19.6GB (fits in 24GB VRAM)
2026-11-17 KRAFTON NVFP4 KRAFTON/Raon-1-mini-preview-nvfp4 –

File sizes of each build:

  • Available builds in KRAFTON/Raon-1-mini-preview-gguf: IQ4_XS 15.0GB / Q4_K_M 18.3GB / Q5_K_M 20.7GB / Q6_K 24.1GB / Q8_0 28.6GB
  • Available builds in KRAFTON/Raon-1-mini-preview-mlx: MLX 16.3GB

In addition, 9 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Other Models for the Same Task

Recent text generation models covered by Local Model Watch, newest first. Grouped by the task each publisher declares on Hugging Face (pipeline_tag); the smallest VRAM tier is this site’s estimate.

See all text generation models →

What to Read Next

Sources

Update History

  • 2026-11-17: Added converted builds to “Quantized and Converted Variants”: KRAFTON/Raon-1-mini-preview-gguf, KRAFTON/Raon-1-mini-preview-mlx, KRAFTON/Raon-1-mini-preview-nvfp4
  • 2026-11-17: Updated the hardware requirements table with the actual file sizes of KRAFTON/Raon-1-mini-preview-gguf.
  • 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).