2026/07/27: 🤗 Released kanana-2-3b, kanana-2-1.3b HF model weights.
2026/07/27: 📕 Published a blog post about the development of the Kanana-2 SLM series.
Introduction
We present Kanana-2 SLM, Kakao's second series of Small Language Models (SLMs), designed to deliver strong language capabilities while remaining compact and efficient for practical deployment. The series includes 3B, 1.3B, and 0.9B models. This release publicly includes the 3B model and the compressed 1.3B model, providing a balance between capability and efficiency for a wide range of applications.
Kanana-2-3B was pretrained from scratch on TPU clusters and further improved through post-training with supervised fine-tuning and reinforcement learning, resulting in strong instruction-following and reasoning capabilities.
Kanana-2-1.3B models are derived from Kanana-2-3B through a cascade pruning and distillation pipeline. To further improve deployment efficiency, they adopt Sliding Window Attention (SWA), enabling memory-efficient long-context inference with support for context lengths of up to 32K tokens while substantially reducing KV-cache memory requirements.
The Kanana-2 SLM release consists of the following four publicly available models:
Kanana-2-3B-Base — 3B pretrained base
Kanana-2-3B-Instruct — instruction-tuned 3B model
Kanana-2-1.3B-Base — compressed 1.3B pretrained base
Kanana-2-1.3B-Instruct — instruction-tuned 1.3B model for on-device deployment
[!NOTE]
No Kakao user data was used for either pre-training or post-training.
Highlights
Cascade Pruning & Distillation: Kanana-2-1.3B is built by progressively compressing Kanana-2-3B-Base (3B → 2B → 1.3B → 0.9B) through a cascade pruning and distillation pipeline.
Sliding Window Attention (SWA): Uses a 3:1 hybrid layout of sliding-window and full-attention layers. A sliding-window size of 1024 reduces per-token KV-cache reads, cutting KV-cache usage by up to ~72.7% at a 32K context length compared to a full-attention-only model. YaRN is applied to full-attention layers, while SWA layers retain RoPE, preserving long-range context without sacrificing local-attention efficiency.
Kanana-2 tokenizer: Improves Korean tokenization efficiency by over 30% compared to the previous generation.
Long context: Natively supports context lengths of up to 32,768 tokens.
Instruction-following, chat, tool-calling, code, math, and knowledge benchmarks for the Kanana-2 SLM series. Scores use greedy decoding (temperature 0.0, top-p 1.0, max 4096 tokens); the metric for each benchmark is listed in the Metric column.
Benchmark
Metric
kanana-2-3b-instruct
kanana-2-1.3b-instruct
Qwen3.5-2B
Qwen3-1.7B
Chat
MT-Bench†
judge
7.15
6.83
6.87
6.98
KoMT-Bench†
judge
6.92
6.54
5.21
5.29
Instruction Following
IFBench
prompt strict
33.33
34.69
24.83
18.33
IFEval
prompt strict
80.96
77.63
66.91
68.39
IHEval
pass@1
35.96
27.06
38.52
42.09
Tool Calling
BFCL-v3 (Live)‡
pass@1
71.94
69.64
66.96
65.48
BFCL-v3 (Multi-Turn)‡
pass@1
17.12
5.50
6.27
4.38
Code Generation
MBPP
pass@1
70.63
69.05
55.56
62.17
MBPP+
pass@1
60.05
60.85
46.83
52.65
Mathematics
GSM-Plus
pass@1
61.17
58.27
61.29
63.10
MATH-500
pass@1
61.20
61.40
67.80
72.00
Minerva Math
pass@1
24.44
22.43
35.18
27.21
Reasoning & Knowledge
MMLU-CoT
acc
61.09
60.37
69.01
66.02
KMMLU-CoT
acc
43.32
42.79
41.75
37.84
HAERAE-Bench (v1.0)-CoT
acc
43.75
44.89
27.84
27.84
KoSimpleQA
acc
22.29
17.81
3.21
2.82
† Evaluated using gpt-4o-2024-08-06 as the judge model.
‡ Live denotes the average score of 6 live benchmarks, and Multi-Turn the average score of 4 multi-turn benchmarks.
Deployment
kanana-2-1.3b-instruct uses a custom hybrid attention architecture (Kanana2TinyForCausalLM — a Qwen3 backbone with a 3:1 SWA/full-attention layout and per-layer-type RoPE), shipped as remote code in the repository.
[!NOTE]
Because the modeling code is loaded from the repository, serving requires transformers >= 4.57 and the --trust-remote-code flag. For SGLang, use --attention-backend triton so the hybrid sliding-window attention is handled correctly.
vLLM
vLLM is a fast and memory-optimized engine designed for high-performance LLM inference and serving.
SGLang is a high-efficiency framework for serving LLMs and VLMs, enabling easy deployment of OpenAI-compatible API servers.
For SGLang, the model is served through the stock Qwen3ForCausalLM path instead of the remote-code Kanana2TinyForCausalLM class. This requires two files shipped in the sglang/ directory of this repository:
sglang/config.json — a Qwen3-flavored config (architectures: ["Qwen3ForCausalLM"], model_type: qwen3, no auto_map) that keeps the hybrid-attention fields (layer_types, sliding_window, per-layer-type rope_parameters). Use it in place of the default config.json when serving with SGLang.
sglang/qwen3.py — a patched model definition that overrides the installed sglang/srt/models/qwen3.py.
With the Qwen3-flavored sglang/config.json the model itself needs no remote code, but --trust-remote-code is passed so the tokenizer/config are loaded without prompting.
Use triton or fa3 for the attention backend. Avoidflashinfer — it appears to have an issue with this model and causes significant accuracy/throughput degradation.
Kanana-2-1.3B shares the Qwen3 backbone but adds a 3:1 SWA/full hybrid attention layout and per-layer-type RoPE, which SGLang's stock Qwen3 model does not handle. The patched sglang/qwen3.py changes the decoder layer to:
Per-layer-type RoPE — for each layer, read the entry in config.rope_parameters matching config.layer_types[layer_id] and build that layer's own rotary embedding from it: full_attention uses YaRN (factor=40, original_max_position_embeddings=4096) and sliding_attention uses default RoPE (rope_theta=10000). A layer type with no matching entry falls back to no RoPE.
Hybrid sliding-window attention — layers typed sliding_attention run RadixAttention with sliding_window_size = config.sliding_window - 1 (SGLang uses an exclusive window, HF an inclusive one), while full_attention layers use full causal attention.
