Raon-1 Mini Benchmarks: How KRAFTON's MoE Model Stacks Up
Raon-1 mini's AIME, GPQA Diamond, and IFBench scores compared against Qwen3.6-35B, Gemma 4 26B, and other MoE models its size.
What is Raon-1 mini and how does it perform on benchmarks?
Raon-1 mini is a 30B-A4B Mixture-of-Experts language model built from scratch by KRAFTON AI, released under Apache 2.0 as a preview. On published benchmark tables, it leads or closely matches similarly sized MoE models like Qwen3.6-35B-A3B and Gemma 4 26B-A4B on reasoning tests (AIME 2026, GPQA Diamond) and instruction-following tasks, while trailing larger dense models such as Qwen3.8-27B on some reasoning scores. It activates only 4.3B of its 28.8B total parameters per token.
TL;DR
- Raon-1 mini is a 30B-A4B MoE model from KRAFTON AI, trained from scratch and released under Apache 2.0 as a preview build, meaning specs can still shift before a final release.
- On AIME 2026, Raon-1 mini scores 93.3, edging out Qwen3.6-35B-A3B (92.9) and Gemma 4 26B-A4B (87.1), though it still falls short of the dense Qwen3.8-27B at 98.8.
- On GPQA Diamond, a graduate-level science reasoning test, Raon-1 mini posts 84.3, ahead of both MoE competitors but again behind Qwen3.8-27B’s 89.7.
- The model uses a hybrid attention architecture mixing 18 linear-attention layers with 7 full-attention layers, which KRAFTON says cuts KV cache size dramatically versus standard transformers.
- Raon-1 mini activates 6 of 128 routed experts plus one shared expert per token, using 4.3B active parameters out of 28.8B total, keeping inference cheap relative to its knowledge base.
- It natively supports 262,144 tokens of context and ships in MXFP8 and NVFP4 quantization, with NVFP4 fitting on a single RTX 5090 (32GB VRAM).
- KRAFTON markets it primarily on Korean-language and agentic task performance, with added training in English, Japanese, Chinese, Spanish, French, and Thai.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
What benchmarks does Raon-1 mini report?
KRAFTON’s model card breaks results into knowledge and reasoning categories, comparing Raon-1 mini against four baselines: two other MoE models (Qwen3.6-35B-A3B, Gemma 4 26B-A4B) and two dense models (Qwen3.8-27B, Muse Glimmer 30B).
On knowledge tests, Raon-1 mini scores 76.5 on MMLU Pro, behind all four comparison models (Qwen3.6-35B-A3B hits 85.9, Qwen3.8-27B hits 85.7). On AA-Omniscience, a benchmark that measures both correctness and model calibration (how often a model knows what it doesn’t know), Raon-1 mini posts 3.5, the only positive score among the five models shown. The others range from -0.5 to -46.6, which suggests Raon-1 mini is comparatively better calibrated even where it hasn’t memorized as much raw factual content.
On reasoning, Raon-1 mini performs more competitively. It scores 84.3 on GPQA Diamond (ahead of both MoE peers, behind both dense models), 93.3 on AIME 2026 (ahead of both MoE peers, behind Qwen3.8-27B’s 98.8), and 84.1 on HMMT Feb 2026 (just behind Qwen3.6-35B-A3B’s 84.5 but well ahead of Gemma 4 26B-A4B’s 79.9). KRAFTON’s own highlight claims it leads comparison models on instruction-following and search benchmarks, though the full IFBench figures weren’t fully legible in the released table.
The pattern that emerges: Raon-1 mini trades some general knowledge breadth for stronger math and science reasoning relative to models its own parameter class, and it’s explicitly positioned as a reasoning-first, Korean-first model rather than a broad knowledge base.
Why does Raon-1 mini beat larger MoE models on some tests?
Part of the answer is architecture. Raon-1 mini uses what KRAFTON calls Raon Delta Attention (RDA), a modification of Kimi Delta Attention that decouples the erase and write gates to improve long-context handling. Eighteen of the model’s 25 layers use this linear-attention mechanism with a fixed-size state, while only 7 layers use full (Gated MLA) attention that requires a traditional KV cache. KRAFTON states this keeps KV cache size around 8 KiB per token in BF16, far smaller than a fully dense-attention model would need at the same context length.
The model also uses Raon-mHC, an efficiency tweak on DeepSeek’s manifold-constrained hyper-connections, which KRAFTON says reduces training compute cost. Combined with a sparse MoE routing scheme (6 routed experts plus 1 shared expert active per token, out of 128 total), the architecture is built to pack more effective reasoning capacity into a smaller activated-parameter footprint.
This is consistent with a broader trend in open-weight models released through 2025 and into 2026: MoE architectures that route to a small fraction of total parameters per token, paired with linear or hybrid attention schemes, have become the default way to push reasoning benchmark scores without proportionally raising inference cost.
How does Raon-1 mini compare to Qwen3.6-35B and Gemma 4 26B specifically?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Against Qwen3.6-35B-A3B, Raon-1 mini is smaller in total and active parameters (28.8B total / 4.3B active versus Qwen’s 35B total / 3B active) yet posts higher scores on AIME 2026 (93.3 vs 92.9) and GPQA Diamond (84.3 vs 83.2). Qwen3.6-35B-A3B wins clearly on MMLU Pro (85.9 vs 76.5) and edges ahead on HMMT Feb 2026 (84.5 vs 84.1).
Against Gemma 4 26B-A4B, which is closer in active-parameter count, Raon-1 mini wins on every reasoning benchmark shown: GPQA Diamond (84.3 vs 80.7), AIME 2026 (93.3 vs 87.1), and HMMT Feb 2026 (84.1 vs 79.9). Gemma 4 26B-A4B does outperform Raon-1 mini on MMLU Pro (82.9 vs 76.5).
The overall shape of the comparison: Raon-1 mini is not the strongest at general knowledge recall among this group, but it’s competitive to leading on multi-step math and science reasoning benchmarks, categories that tend to track more closely with real-world coding and agentic task performance than trivia-style knowledge tests do.
Is Raon-1 mini worth using for coding or agentic tasks?
The model card emphasizes agent execution as a core feature rather than an afterthought. Raon-1 mini supports tool calling for agentic workflows and lists compatibility with Hermes Agent, Pi, and OpenCode. KRAFTON also ships a DFlash2 draft model specifically for speculative decoding, a technique that speeds up generation by having a smaller model draft tokens that the main model verifies, which matters for latency-sensitive coding assistants and agent loops.
The long context window (262,144 tokens natively) also matters for agentic use cases that require holding large codebases, tool outputs, or multi-turn histories in context without aggressive truncation.
That said, this is a preview release. KRAFTON states explicitly that specs and results may change before a final version ships, so teams building production pipelines around Raon-1 mini should treat current benchmark numbers as provisional rather than locked in.
Frequently Asked Questions
What does “30B-A4B” mean for Raon-1 mini?
It describes the MoE architecture: roughly 30 billion total parameters (28.8B specifically), with only about 4.3 billion activated per token through sparse expert routing. This keeps inference costs closer to a 4B dense model while retaining the knowledge capacity of a much larger one.
Can Raon-1 mini run on consumer hardware?
Yes, in its NVFP4 quantized format. KRAFTON states this quantization enables inference on a single NVIDIA RTX 5090 with 32GB of VRAM. An MXFP8 format is also available for setups with more headroom.
How does Raon-1 mini’s context window compare to other open models?
It natively supports up to 262,144 tokens, which is in line with or ahead of many open-weight models in its size class as of this release, aided by its hybrid attention design that limits KV cache growth.
Is Raon-1 mini good at languages other than Korean?
It’s trained on Korean, English, Japanese, Chinese, Spanish, French, and Thai, with KRAFTON describing Korean performance as strong and English capability as solid. It’s explicitly positioned as a Korean-first model with broader multilingual support layered on.
Why does Raon-1 mini score lower on MMLU Pro but higher on AIME and GPQA Diamond?
MMLU Pro tests broad factual and professional knowledge recall, while AIME and GPQA Diamond test multi-step mathematical and scientific reasoning. Raon-1 mini’s training and architecture appear tuned toward reasoning performance specifically, trading some raw knowledge breadth for stronger step-by-step problem solving.