Raon-1 nano

KRAFTON's 3.1B on-device model (1.5B effective), distilled from Raon-1 mini preview, for Korean and English with tool calling.

In short. Raon-1 nano is a small text model from KRAFTON, the South Korean game company, made to run on a laptop or a phone. It follows instructions, keeps track of a conversation and calls tools in Korean and English, and it was shrunk from the larger Raon-1 mini preview. The 4-bit builds are only about 2 GB. It is released under Apache 2.0.

Summary

Raon-1 nano is a dense language model from KRAFTON AI with 3.1B parameters in total, of which 1.5B are an embedding table that is read per layer, so the card calls it a 1.5B-effective model. Its Hugging Face repository was created on 29 October 2026 and carries an Apache 2.0 notice added on 9 November, and the GGUF, MLX and LiteRT-LM conversions followed on 10 November. KRAFTON says it was derived from the 30B-A4B Raon-1 mini preview by MoE-to-Dense pruning and knowledge distillation, then trained on a further 3T tokens, fine-tuned, distilled on-policy from the mini model and finished with reinforcement learning. It is built to follow instructions, keep conversational context and call tools in Korean and English, for tasks such as office work and controlling a phone.

The attention layout matches its larger sibling: 18 of 24 layers are Raon Delta Attention layers with a fixed-size state, and only the 6 MLA layers keep a KV cache, about 7 KiB per token in BF16, with a native context of 262,144 tokens. A 55M-parameter multi-token-prediction module ships in the checkpoint. The BF16 repository is 7.2 GB. KRAFTON also publishes GGUF files (2.1 GB at Q4_K_M and 3.4 GB at Q8_0), a 4-bit MLX model of about 1.8 GB for Apple silicon, and a 4-bit LiteRT-LM bundle of about 1.8 GB that it checked on Apple silicon GPUs; that bundle has a 4,096-token context and answers without thinking.

In KRAFTON's table it is compared with Qwen3.5 2B, Gemma 4 E2B, MiniCPM5 2B and LFM3 1.7B. It scores highest on instruction following (IFBench 75.9, against 73.7 for LFM3), grounding (FACTS Grounding 44.4), personalization (PersonaMem v2 38.8, PrefEval 44.0), mobile control (AndroidControl 36.0) and the Korean tests: KMMLU-Redux 50.9, CLIcK 65.2 and an internal Korean chat benchmark where it scores 91.0 against 87.6 for Gemma 4 E2B. It trails on general knowledge (MMLU-Redux 81.1 against 84.0 for MiniCPM5 2B), function calling (BFCL v4 56.3 against 67.2 for LFM3 1.7B and 61.0 for MiniCPM5 2B), terminal commands (NL2SH-ALFA 58.0 against 62.6 for Gemma 4 E2B) and search (BrowseComp Plus 34.5 against 37.2 for MiniCPM5 2B). The office-work and Korean-chat benchmarks are internal and unreleased.

Like the mini model, it is designed for thinking mode, and the card recommends a temperature of 0.1 with a presence penalty of 1.5 for general use and 1.0 with no penalty for agent and code work. The repository includes a LICENSE file; the Apache 2.0 grant covers the weights and repository files, not KRAFTON's training data or other proprietary materials.

Good for

  • Korean and English chat, instruction following and grounded answers on a laptop, a Mac or a phone
  • Tool calling for everyday tasks such as office work and mobile control, where it leads the compared small models on AndroidControl
  • Apple silicon and mobile runtimes, with MLX and LiteRT-LM bundles of about 1.8 GB
  • A lighter companion to Raon-1 mini preview when you want the same Korean-first behavior on much smaller hardware

Watch out for

  • All scores are KRAFTON's own, and two of the benchmarks are internal and unreleased
  • It trails the small models it is compared with on function calling (BFCL v4), terminal commands and search
  • The LiteRT-LM bundle has only a 4,096-token context and does not think, unlike the main model
  • The files hold all 3.1B parameters, including the 1.5B embedding table, so plan memory from the file sizes rather than from the 1.5B effective figure
  • It is brand new, with little independent testing so far

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the MLX (4-bit GPTQ) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Raon-1-nano (BF16) BF16 3.1B 7.13 GB 3.56 GB 2.14 GB Laptop or phone
GGUF Q8_0 INT8 3.1B 7.13 GB 3.56 GB 2.14 GB Laptop or phone
GGUF Q4_K_M INT4 3.1B 7.13 GB 3.56 GB 2.31 GB Laptop or phone
MLX (4-bit GPTQ) INT4 3.1B 7.13 GB 3.56 GB 1.98 GB Laptop or phone
LiteRT-LM bundle (4-bit) INT4 3.1B 7.13 GB 3.56 GB 1.98 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Raon-1 nano?
Raon-1 nano is a small text model from KRAFTON, the South Korean game company, made to run on a laptop or a phone. It follows instructions, keeps track of a conversation and calls tools in Korean and English, and it was shrunk from the larger Raon-1 mini preview. The 4-bit builds are only about 2 GB. It is released under Apache 2.0.
Can I use Raon-1 nano commercially?
Yes. Raon-1 nano is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Raon-1 nano open source?
Raon-1 nano is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Raon-1 nano on my own computer?
Yes. The smallest variant, MLX (4-bit GPTQ), needs about 1.98 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Raon-1 nano released?
Raon-1 nano was released by KRAFTON on November 2026.
How long a context window does Raon-1 nano have?
Raon-1 nano supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Raon-1 mini preview KRAFTON 28.8B 4.3B active 256K Apache 2.0 (Permissive) 18 GB Gaming GPU
LFM3 Liquid AI 1.73B – 2.17B 256K LFM Open 1.0 (Conditional) 1.19 GB Laptop or phone
MiniCPM5-2B OpenBMB 1.08B – 2.52B 128K Apache 2.0 (Permissive) 0.745 GB Laptop or phone
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone