--- title: "How Does Raon-1 mini Score vs Qwen3.6 and Gemma 4?" url: "https://goremy.ai/blog/articles/raon-1-mini-benchmarks-vs-qwen-gemma-4" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-10-23T13:40:06.802Z" updated: "2026-10-23T13:40:06.802Z" topic: AI Model Reviews & Comparisons description: Raon-1 mini preview benchmark table against Qwen3.6-35B-A3B, Gemma 4 26B-A4B and dense peers on GPQA Diamond, AIME 2026, MMLU Pro and more. --- # How Does Raon-1 mini Score vs Qwen3.6 and Gemma 4? > Raon-1 mini preview benchmark table against Qwen3.6-35B-A3B, Gemma 4 26B-A4B and dense peers on GPQA Diamond, AIME 2026, MMLU Pro and more. Edited by Luis Chavez-Mattos, Director, Product · Published 2026-10-23 ## How does Raon-1 mini preview score against Qwen3.6 and Gemma 4? Raon-1 mini preview, a 30B-A4B mixture-of-experts model from KRAFTON AI, beats both MoE rivals on GPQA Diamond (84.3) and AIME 2026 (93.3) and posts the only positive AA-Omniscience score in the comparison (3.5). It trails Qwen3.6-35B-A3B and Gemma 4 26B-A4B on MMLU Pro (76.5 vs 85.9 and 82.9). The numbers come from KRAFTON's own model card, so treat them as vendor-reported. ## TL;DR - Raon-1 mini preview scores **84.3 on GPQA Diamond**, ahead of Qwen3.6-35B-A3B (83.2) and Gemma 4 26B-A4B (80.7). - On **AIME 2026 it hits 93.3**, a thin edge over Qwen3.6 (92.9) and a clear gap over Gemma 4 (87.1). - **MMLU Pro is the weak spot** at 76.5, roughly 6 to 9 points behind the MoE peers and the dense models. - It is the only model in the table with a **positive AA-Omniscience score** (3.5), while the others land between −0.5 and −46.6. - The dense **Qwen3.8-27B beats it on GPQA Diamond and AIME 2026**, so the wins are among MoE peers, not across the board. - The release is a **preview under Apache 2.0**, and KRAFTON says specs and results may change before the final version. ## What is Raon-1 mini? Raon-1 mini is an open-weight model trained from scratch by KRAFTON AI and released under the Apache 2.0 license. It has 28.8B total parameters and activates 4.3B per token. Each token routes through 6 of 128 experts plus 1 shared expert. KRAFTON positions it for Korean conversation and agentic tasks, with built-in thinking. The architecture is a hybrid. Of its 25 layers, 18 use Raon Delta Attention (a linear attention variant that decouples the erase and write gates of Kimi Delta Attention). The other 7 are gated MLA layers. Only those 7 keep a KV cache, about 8 KiB per token in BF16. Native context is 262,144 tokens. ## How do the benchmark numbers compare? The model card compares Raon-1 mini preview against two MoE models (Qwen3.6-35B-A3B and Gemma 4 26B-A4B) and two dense models (Qwen3.8-27B and Muse Glimmer 30B). | Benchmark | Raon-1 mini preview (30B-A4B) | Qwen3.6-35B-A3B | Gemma 4 26B-A4B | Qwen3.8-27B (dense) | Muse Glimmer 30B (dense) | |---|---|---|---|---|---| | MMLU Pro | 76.5 | 85.9 | 82.9 | 85.7 | 84.0 | | AA-Omniscience | 3.5 | −7.2 | −46.6 | −0.5 | −34.7 | | GPQA Diamond | 84.3 | 83.2 | 80.7 | 89.7 | 82.6 | | AIME 2026 | 93.3 | 92.9 | 87.1 | 98.8 | 92.9 | | HMMT Feb 2026 | 84.1 | 84.5 | 79.9 | n/a | n/a | The HMMT row is incomplete because the dense columns were cut off in the material I had. The card also covers instruction-following and search benchmarks, where KRAFTON says Raon-1 mini leads the comparison models. I could not verify those exact scores, so check the model card for the figures. ## Where does Raon-1 mini win? Against the MoE peers, the pattern is clear on reasoning. - **GPQA Diamond:** 84.3 vs 83.2 for Qwen3.6 and 80.7 for Gemma 4. It also edges the dense Muse Glimmer 30B (82.6). - **AIME 2026:** 93.3, just ahead of Qwen3.6 (92.9) and well ahead of Gemma 4 (87.1). It matches or slightly beats Muse Glimmer 30B (92.9). - **AA-Omniscience:** 3.5, the only positive score in the table. This benchmark penalizes wrong answers, so a higher number suggests the model is better at not guessing when it does not know. The gaps are large. Gemma 4 sits at −46.6. The AIME margin over Qwen3.6 is 0.4 points. On a benchmark with only a few dozen problems, that is within noise. The GPQA margin of about one point is also small. The Gemma 4 gaps are the more meaningful ones. ## Where does it fall short? MMLU Pro is the clear miss. At 76.5, Raon-1 mini sits behind Qwen3.6 (85.9), Gemma 4 (82.9), Qwen3.8-27B (85.7) and Muse Glimmer (84.0). Broad knowledge recall is where a 4.3B-active model trained from scratch would be expected to show its limits, and this number reflects that. On HMMT Feb 2026 it lands at 84.1, a hair under Qwen3.6 (84.5) and above Gemma 4 (79.9). The dense Qwen3.8-27B is stronger on the reasoning tests shown: 89.7 on GPQA Diamond and 98.8 on AIME 2026. The model card's highlighting marks Raon's best results among MoE baselines, so the dense columns are context, not a contest it wins. ## Is the Korean-language angle visible in these benchmarks? Not in the numbers I could see. The table covers knowledge, reasoning and (per the card) instruction following and search. KRAFTON states that the model was trained on Korean, English, Japanese, Chinese, Spanish, French and Thai, and that it shows strong Korean performance and solid English. Those claims stand on the card's text, not on a Korean benchmark row I can quote. If Korean quality matters to you, test it on your own prompts. ## What does it take to run Raon-1 mini? The card lists several deployment details: - **Quantization:** MXFP8 and NVFP4 formats are available. - **Single-GPU inference:** NVFP4 enables inference on one NVIDIA RTX 5090 (32 GB VRAM). - **Speculative decoding:** a DFlash2 draft model is provided. - **Agent tooling:** tool calling is supported, with named support for Hermes Agent, Pi and OpenCode. - **License:** Apache 2.0. The small KV cache from the hybrid design also matters for long-context work, since only 7 of 25 layers store per-token state. ## Should you trust these benchmarks? Use some caution. These are published by the model's developer, the model is a preview, and I have no independent reproduction. Evaluation settings such as thinking budget, sampling parameters and prompt format can move scores by several points. A fair read: Raon-1 mini looks competitive on math and science reasoning for its active size, and weaker on broad knowledge. Run your own evals on the tasks you care about before switching anything. ## Frequently Asked Questions ### Is Raon-1 mini better than Qwen3.6-35B-A3B? It depends on the task. Raon-1 mini preview scores higher on GPQA Diamond (84.3 vs 83.2), AIME 2026 (93.3 vs 92.9) and AA-Omniscience. Qwen3.6 is stronger on MMLU Pro (85.9 vs 76.5) and slightly ahead on HMMT Feb 2026 (84.5 vs 84.1). ### How does Raon-1 mini compare with Gemma 4 26B-A4B? Raon-1 mini leads Gemma 4 on every row in the table I could see: GPQA Diamond, AIME 2026, HMMT Feb 2026 and AA-Omniscience, except MMLU Pro, where Gemma 4 scores 82.9 against 76.5. The AA-Omniscience gap is the largest, at 3.5 vs −46.6. ### Is Raon-1 mini open source? It is open-weight and released under the Apache 2.0 license. KRAFTON describes it as part of the Raon open model family. It is a preview, so the final release may differ in specs and results. ### Can I run Raon-1 mini on a single consumer GPU? According to the model card, the NVFP4 quantized version can run inference on a single NVIDIA RTX 5090 with 32 GB of VRAM. MXFP8 is also offered. Actual speed and context limits will depend on your inference stack.