Model reference · open weights

kanana-1.5-v

LLMs kakaocorp Vision + text 1 build Its own licence terms 28k dl/mo

kanana-1.5-v is an open-weight language model from kakaocorp. kanana-1.5-v-3b-instruct (BF16) weighs 7.3 GB; the smallest configuration that runs it is RTX 3060 12 GB.

Request Rent a GPU

For self-hosting, compare this build's requirements below with the GPU you plan to rent.

kanana-1.5-v-3b-instruct is a multimodal large language model developed by Kakao that accepts text and image inputs to generate text.
It is designed for tasks such as image captioning, document understanding, and multimodal instruction following in English and Korean.
The model contains 3.7B parameters, supports a context length of 32,768 tokens, and is released under the kanana-license.

Summary of the kakaocorp/kanana-1.5-v-3b-instruct model card, 2026-10-01

Hugging Face ↗

What it is

Released by	kakaocorp
Released	2025-07-23
Parameters	3.7B
VRAM	7.3 GB for the weights

What it runs on

Memory and cards for kanana-1.5-v-3b-instruct (BF16)

7.3 GBweights, file size

131 MBcache per 1K tokens

2.0 GBruntime overhead, at least

32,768 tokenscontext max

Card	Requests at once		Context max	Memory
Card	8K each	32K each	Context max	Memory
RTX 3060 12 GB	2		17K	11.6 GB
RTX 4060 Ti 16 GB	5	1	all 32K	15.4 GB
RTX 3090 24 GB	13	3	all 32K	23.4 GB
RTX 4090 24 GB	13	3	all 32K	23.4 GB
RTX 5090 32 GB	20	5	all 32K	31.0 GB
L40S 48 GB	32	8	all 32K	44.0 GB
A100 80 GB	64	16	all 32K	78.2 GB
H100 80 GB	60	15	all 32K	78.1 GB
RTX PRO 6000 Blackwell 96 GB	75	18	all 32K	93.8 GB
DGX Spark (GB10) 128 GB unified	87	21	all 32K	107 GB
H200 141 GB	116	29	all 32K	138 GB
B200 180 GB	151	37	all 32K	176 GB
2× RTX 3060 12 GB tensor parallel	11	2	all 32K	11.6 GB a card
2× RTX 4060 Ti 16 GB tensor parallel	18	4	all 32K	15.4 GB a card

Memory needed at each load

Requests at once	8K tokens each	32K tokens each
1	10.4 GB	13.6 GB
5	14.7 GB	30.8 GB
8	17.9 GB	43.7 GB
16	26.5 GB	78.0 GB
32	43.7 GB	147 GB
64	78.0 GB	284 GB

One card, with vLLM's small-card settings.

From the model card

What kakaocorp says about kanana-1.5-v

Read the model card

🤗 1.5 HF Models &nbsp | &nbsp 📕 Blog &nbsp

kanana-1.5-v-3b-instruct

The Unified Foundation Model (UFO) task force of Kanana at Kakao developed and released the Kanana-V family of multimodal large language models (MLLMs), a collection of pretrained text/image-to-text (TI2T) models.

Intended Use

kanana-1.5-v-3b-instruct is intended for research and application development in multimodal understanding and text generation tasks. Typical use cases include image captioning, document understanding, OCR-based reasoning, and multimodal instruction following in both English and Korean. The model is optimized for both general-purpose and Korea-specific benchmarks, making it suitable for bilingual environments.

Model Details

Developed by: Unified Foundation Model (UFO) TF at Kakao
Language(s) : ['en', 'ko']
Model Architecture: kanana-1.5-v-3b-instruct has 3.6B parameters and contains image encoder, C-abstractor, and kanana-1.5-3b-instruct language model.
Input: The models accept text and image inputs.
Output: The models generate text only.
Context Length: 32k
Knowledge Cutoff Date: June 30, 2024
Model Release Date: Jul 24, 2025.
License: kanana-license

Evaluation

Model Configuration Summary

Model	LLM	Total Parameter
kanana-1.5-v-3b-instruct	kanana-1.5-3b-instruct	3.67B
HCX-SEED-Vision-3B	HyperCLOVAX-SEED-Text-Base-3B	3.72B
Phi-3-Vision	Phi-3-Mini	4.15B
Qwen2.5-VL-3B-Instruct	Qwen2.5-3B	3.75B
InternVL2.5-4B	Qwen2.5-3B-Instruct	3.94B

Overview

Model	All	Image (EN)	Image (KO)	IF (EN, KO)
kanana-1.5-v-3b-instruct	73.22	74.00	68.27	77.39
HCX-SEED-Vision-3B	59.00	64.81	51.96	60.23
Phi-3-Vision	48.84	65.41	36.40	44.71
Qwen2.5-VL-3B-Instruct	63.54	73.97	60.60	56.04
InternVL2.5-4B	61.35	74.73	54.68	54.63

Image Benchmarks (EN)

Model	average	MMMU (Val)	MathVista	DocVQA	ChartQA	OCRBench	InfoVQA	TextVQA	RealWorldQA	MMStar	MMB	SEED-image	MMVet	LLaVA-Wild	scienceqa	AI2D
kanana-1.5-v-3b-instruct	74.00	43.89	56.00	93.06	81.20	82.50	73.62	78.62	65.36	56.32	78.44	75.17	65.87	89.60	95.61	74.81
HCX-SEED-Vision-3B	64.81	38.89	47.40	79.87	71.88	62.90	55.59	73.51	62.48	46.66	72.42	74.84	47.27	79.30	86.84	72.31
Phi-3-Vision	65.41	45.33	43.60	87.04	81.40	63.60	54.80	69.61	59.08	47.47	73.37	71.69	45.96	70.40	90.84	76.98
Qwen2.5-VL-3B-Instruct	73.97	50.67	62.00	94.19	83.60	79.10	77.22	77.77	59.74	56.26	77.75	74.83	61.06	96.90	79.69	78.79
InternVL2.5-4B	74.73	52.33	61.80	92.13	82.76	79.20	69.73	78.24	62.88	59.72	81.96	75.59	61.38	86.30	97.14	79.83

Image Benchmarks (KO)

Model	average	KoOCRBench	KoMMDBench	KoChartTask	KoMathSolution	KoCosMed	KoFoodMenu	KoEntity	KoExam	KoCelebV2
kanana-1.5-v-3b-instruct	68.27	85.93	74.00	84.96	36.88	87.58	70.84	72.04	58.99	43.24
HCX-SEED-Vision-3B	51.96	32.91	64.57	73.55	27.88	78.16	57.08	64.12	31.82	37.58
Phi-3-Vision	36.40	25.13	37.93	52.36	38.75	56.75	34.70	31.71	24.05	26.25
Qwen2.5-VL-3B-Instruct	60.60	50.67	61.75	84.96	47.13	82.01	66.32	58.15	60.68	33.72
InternVL2.5-4B	54.68	20.52	62.65	82.61	46.50	82.66	65.09	50.42	47.43	34.23

Multimodal Instruction Following Benchmarks (EN, KO)

| Model | average | MIABench | MIABench-Ko | MM-IFEval | MM-OmniAlign | |----------------------------|--

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

Details

Type	Language models
Task	Vision + text
Context	32,768 tokens
Runs with	transformers
Based on	kakaocorp/kanana-1.5-v-3b-instruct
Popularity	28k downloads / month
Weights	7.3 GB, kanana-1.5-v-3b-instruct (BF16, file size)

Languages

koen