Kanana-2-v-decide-3b-2611

Kanana-2-v-decide-3b-2611 is Kakao's 3.67B parameter open-weight multimodal decision model. Released under the commercial-friendly Kanana Open License, it introduces an alternative architecture to standard autoregressive text generators. Rather than generating sequences of text tokens to answer prompts, Kanana-2-v-decide evaluates an input state (text, JSON, chat history, or up to 10 images) alongside a set of candidate answers, returning calibrated probability distributions, categorical choices, binary decisions, or numerical scores.

This architecture directly addresses the operational bottlenecks of local agentic AI pipelines. In multi-step agent frameworks, local models are frequently called simply to classify intent, route execution paths, select tools, or verify visual states. Using standard causal language models for these tasks wastes compute cycles on token decoding and introduces formatting errors or hallucinations. Kanana-2-v-decide functions as an ultra-fast classification and scoring engine, executing a single forward pass that outputs precise probabilities rather than free-form text.

With an active parameter footprint of 3.67 billion parameters and support for up to 20,480 tokens of context, the model fits cleanly into consumer GPU setups. For developers who want to run Kanana-2-v-decide-3b-2611 locally, it offers a specialized, low-latency utility layer that outperforms general-purpose small language models on structured decision-making tasks, particularly in Korean and English environments.

Kanana-2-v-decide-3b-2611 is a dense multimodal network built on Kakao's Kanana-2-v-3b foundation. The total 3.67B parameter budget is split into two specialized components: a 0.49B parameter vision encoder paired with a 3.18B parameter language model backbone.

Unlike conventional vision-language models (VLMs) that feed image tokens into an autoregressive decoder to generate descriptive text, Kanana-2-v-decide operates as a dual-encoder-style scoring engine using a shared representation space:

• Embedding Space: The model produces 2,560-dimensional embeddings.
• State and Question Encoding: The input state (system prompts, dialogue history, raw text, or JSON payloads) and the target question are concatenated and projected into a single unified context embedding. If images are present, up to 10 image inputs are processed via the 0.49B vision encoder and integrated into this representation.
• Candidate Projection: Each candidate answer option is independently mapped to an embedding within the same 2,560-dimensional space.
• Scoring via Cosine Similarity: The model computes the cosine similarity between the state-question vector and each candidate vector. A softmax function over these similarity values yields a fully calibrated probability distribution across the options.

This dense design ensures uniform execution latency regardless of the number of options evaluated. Because the network processes up to 20,480 tokens of context, it can evaluate extensive chat logs or deeply nested JSON schemas alongside high-resolution imagery without aggressive truncation. The absence of autoregressive token generation eliminates key-value (KV) cache expansion during output generation, keeping memory footprints static during inference.

Kanana-2-v-decide-3b-2611 is engineered specifically for zero-shot decision, classification, and scoring pipelines. It supports three distinct query formats:

• Choice: Multi-class classification selecting the best fit from an arbitrary list of candidate options.
• Noul: Binary classification (yes/no or true/false verification).
• Score: Regression-style evaluation over an ordered or graded numerical scale.

On Kakao's Korean evaluation benchmark (Kanana Decision Bench), the model achieves a score of 72.50 across 15,909 text questions, outperforming competing small models by wide margins. On standard international evaluations, it scores 61.34 on Decision Index and 61.12 on Decision Index Vision (averaging 61.23 across public benchmarks).

• Agent Routing and Tool Selection: In local agent loops, routing user requests to the correct Python function or API endpoint is a frequent requirement. Kanana-2-v-decide can evaluate tool descriptions against user context in a single forward pass, providing the chosen tool alongside a confidence score.
• Visual Defect and State Inspection: With support for up to 10 images, the model handles physical verification tasks. E-commerce platforms can pass product photos alongside customer return requests to determine if goods are damaged, missing parts, or incorrectly categorized.
• Bilingual Content Moderation: Operating natively across Korean and English, the model categorizes incoming customer inquiries, flags policy violations, or evaluates message urgency directly inside corporate networks.
• RAG Document Re-ranking: By setting candidate passages against query criteria and utilizing the score evaluation mode, the model functions as a cross-encoder ranker across 20,480 tokens of retrieved context.

Running this model locally requires understanding its execution profile. Because it does not generate text token by token, standard metrics like Kanana-2-v-decide-3b-2611 tokens per second do not apply. Instead, Kanana-2-v-decide-3b-2611 performance is measured by end-to-end forward-pass latency. On datacenter silicon such as the Nvidia H200, Kakao reports response times under 0.04 seconds. On consumer hardware, latency remains low because inference consists of a single prompt-processing phase followed by vector similarity calculations.

At 3.67B parameters, the model is exceptionally light on memory, making it accessible on single consumer GPUs, mini PCs, and unified-memory laptops.

If pushing the context window toward its 20,480-token maximum with multiple images loaded, allocate an additional 2 GB to 4 GB of VRAM for the vision encoder activations and sequence attention buffers.

• Best GPU for Kanana-2-v-decide-3b-2611: An Nvidia GeForce RTX 4060 (8GB) or RTX 3060 (12GB) easily runs the model in native BF16 precision. For multi-agent frameworks requiring high concurrent throughput across dozens of parallel routing calls, an RTX 4090 or RTX 4080 allows batch evaluation with sub-20ms latency.
• Apple Silicon: Macs equipped with an M2, M3, or M4 chip (minimum 16GB unified memory) run the model comfortably through PyTorch or MLX backends, processing multimodal inputs without memory pressure.
• Edge Compute: Single-board devices such as the Nvidia Jetson Orin (16GB) can host this local AI model 3.67B parameters 2026 release directly on edge robotics or industrial sorting lines.

For the vast majority of local deployments, the best quantization for Kanana-2-v-decide-3b-2611 is Q8_0 or unquantized BF16. Because the model relies on fine-grained cosine similarity differences across 2,560-dimensional embeddings to calibrate output probabilities, aggressive 4-bit quantization (such as Q4_K_M) can cause minor probability drift in edge-case classification choices. Given that unquantized BF16 fits entirely within an 8GB to 12GB GPU frame, running at 8-bit or 16-bit precision is strongly recommended unless constrained by ultra-low-memory mobile hardware.

Because Kanana-2-v-decide uses a specialized decision head (system_one) rather than standard causal LM generate loops, it currently requires deployment through Hugging Face transformers with custom code execution enabled:

1import torch
2from transformers import AutoModel
3
4model = AutoModel.from_pretrained(
5    "kakaocorp/kanana-2-v-decide-3b-2611", 
6    trust_remote_code=True,
7    torch_dtype=torch.bfloat16
8).to("cuda")
9
10result = model.system_one(
11    state="Customer received broken ceramic plate.",
12    questions={
13        "category": {
14            "type": "choice",
15            "instructions": "Determine support department.",
16            "criteria": {
17                "shipping": "Damaged during transit",
18                "billing": "Pricing or payment issue",
19                "general": "General questions"
20            }
21        }
22    }
23)

While community runtimes like Ollama and llama.cpp focus predominantly on autoregressive text generation, custom embeddings-based inference servers (or vLLM pull requests supporting classification architectures) are the primary vehicle for serving this model in production.

To evaluate whether to run Kanana-2-v-decide-3b-2611 locally, compare it against competing decision and routing architectures in the 2B to 3B parameter range.

StartLux-Decision-v1.1-2B is a dedicated 2B parameter decision checkpoint optimized for rapid English routing. While StartLux features slightly lower memory consumption due to its smaller footprint, Kanana-2-v-decide-3b-2611 holds a decisive advantage in multimodal workflows and multilingual accuracy. Kanana leads on the public Decision Index benchmark (61.34 vs 59.17) and significantly dominates Korean language tasks, scoring 72.50 on Kanana Decision Bench compared to StartLux's 54.38.

Liquid AI's d1-3B represents a non-transformer alternative in the 3B weight class designed for system-one decision operations. While Liquid's architecture offers competitive memory handling on sustained sequence streaming, Kanana-2-v-decide outperforms it on benchmarked decision accuracy. Across public evaluations, Kanana scores an average of 61.23 across text and vision decision indices, while Liquid AI d1-3B registers a 50.42 average. Kanana also natively integrates multi-image processing (up to 10 visual inputs), whereas d1-3B is primarily constrained to text states.

The primary tradeoff when selecting Kanana-2-v-decide-3b-2611 is its operational scope. It cannot draft natural language responses, write code, or explain its reasoning to an end user. It requires the developer to formulate problems as discrete decisions, scoring matrices, or categorical choices. However, when deployed upstream as a router or safety classifier alongside larger generative models, it delivers higher reliability, calibrated confidence scores, and dramatically lower latency than general-purpose text LLMs.