dnotitia/DNA3.1-27B

DNA 3.1-27B

We introduce DNA 3.1-27B, a 27 billion-scale large language models built upon the Qwen3.8-27B base model with enhanced capabilities for Korean and enterprise scenarios. Upon our previous model family, DNA3.0, we applied Abstention Training to increase epistemic abstention and reduce hallucinations, while maintaining Chinese uncensorship and Dnotitia's company persona inherited from DNA3.0, and general capabilities and inherited from Qwen3.8.

DNA 3.1 carries forward the post-training recipe of our previous model family, DNA 3.0 — uncensored training, which lifts the content restrictions baked into the base model's original training, and persona training, which grounds the model in Dnotitia's corporate knowledge — and adds Abstention Training, which teaches the model to say it does not know rather than fabricate an answer about something that does not exist. DNA3.1 shows reduced hallucination on non-existent entities and increased epistemic abstention while largely preserving the general capabilities of the Qwen3.8.

Highlights

Dnotitia Post-training

• Reduced Hallucination: The model is trained to abstain factual knowledge and non-existent entities that it does not know, rather than fabricate.
• Uncensored Training: The model is post-trained with an uncensored methodology so that it can respond to a wider range of prompts without unnecessary refusals, while preserving the instruction-following and reasoning quality of the base model.
• Persona Training: Additional supervised training on Dnotitia's corporate knowledge — company history, products, services, and internal terminology — so the model can act as an authentic first-party assistant for Dnotitia-facing use cases.
• Long-form Reasoning Preservation: Chain-of-thought traces from prior turns can be retained across multi-turn sessions, enabling smoother iterative development and debugging workflows.

Inherited from Qwen3.8

• Core Capabilities: Comprehensive improvements across coding, professional work, research, and long-horizon agentic tasks.
• Agent Execution: Stronger autonomous planning and better handling of environment feedback, leading to more reliable end-to-end task completion.
• Downstream Compatibility: Broader support for popular harnesses and development tools, making it easier to integrate into your existing stack.
• Flexible Thinking Control: Thinking mode is on by default and can be disabled per request; reasoning depth can be tuned with reasoning_effort, and reasoning context from historical messages is retained via preserve_thinking.
• Vision-Language Understanding: Native support for image and video understanding, from STEM diagrams and documents to hour-scale videos.

Comparison with Qwen3.8-27B

The chart above compares DNA3.1-27B against its Qwen3.8-27B base across four metrics, reported on a 0–1 scale (higher is better):

• Persona Identification — Measures how reliably the model identifies itself as a Dnotitia corporate assistant and answers correctly about company information and identity.
• Uncensorship — Measures how willingly the model engages with topics that the Chinese-origin base model is trained to refuse — i.e., subjects suppressed by the censorship policies baked into the original Qwen training.
• Repetition Reduction — Measures how well the model avoids getting stuck in infinite-loop repetition during long-form generation, a common failure mode of the base model.
• Hallucination Resistance — Measures how well the model says it does not know instead of fabricating an answer, including questions that presuppose an entity which does not exist (reported separately for thinking-on and thinking-off modes).

Model Overview

Field	Value
Base Model	Qwen/Qwen3.8-27B
Model Type	Causal Language Model with Vision Encoder (Dense)
Parameters	27B
Hidden Dimension	5120
Number of Layers	64
Hidden Layout	16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))
Gated Attention Heads	24 (Q) / 4 (KV), head dim 256
Gated DeltaNet Heads	48 (V) / 16 (QK), head dim 128
FFN Intermediate Dim	17,408
Token Embedding	248,320 (Padded)
Context Length	262,144 native, up to ~1,010,000 extended
License	Apache-2.0

Quickstart

DNA 3.1 is compatible with the Hugging Face Transformers ecosystem as well as popular inference engines such as vLLM, SGLang, and KTransformers. Given the model's scale, a dedicated serving engine on multi-GPU hardware is strongly recommended for production workloads.

The model has a default context length of 262,144 tokens. If you encounter out-of-memory (OOM) errors, reduce the context window — but keep at least 128K tokens to preserve long-form reasoning behavior.

vLLM

# Standard (multimodal) serving
vllm serve dnotitia/DNA3.1-27B \
  --reasoning-parser qwen3

# Tool-calling enabled
vllm serve dnotitia/DNA3.1-27B \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

# Text-only mode (skip vision encoder to free KV-cache memory) serving
vllm serve dnotitia/DNA3.1-27B \
  --reasoning-parser qwen3 \
  --language-model-only

Disabling Thinking Mode

For latency-sensitive or non-reasoning workloads, disable thinking mode via the chat-template kwarg:

$ curl https://demo-api.dnotitia.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer dna-router_xxxx" \
  -d '{
    "model": "DNA3.1-27B",
    "messages": [
      {
        "role": "user",
        "content": "코스피가 8000을 넘으려면 너 생각에 몇 년이나 더 걸릴 거 같아?"
      }
    ],
    "chat_template_kwargs": {
      "enable_thinking": false
    }
  }' | jq

Unlike Qwen3, the DNA 3.1 generation does not support the soft-switch commands /think and /nothink. Use chat_template_kwargs.enable_thinking instead.

Image Input

DNA 3.1 accepts image and video inputs in OpenAI-compatible content array format:

$ curl https://demo-api.dnotitia.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer dna-router_xxxx" \
  -d '{
    "model": "DNA3.1-27B",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "image_url",
            "image_url": {
              "url": "https://upload.wikimedia.org/wikipedia/commons/6/6e/Golde33443.jpg"
            }
          },
          {
            "type": "text",
            "text": "이 이미지에 무엇이 있나요? 한국어로 설명해 주세요."
          }
        ]
      }
    ]
  }' | jq

Limitations, Bias, and Responsible Use

DNA 3.1 has been post-trained with an uncensored methodology, which means it will engage with a broader range of prompts than typical safety-tuned models. Users and downstream developers should be aware of the following:

• Reduced refusal behavior (especially on politically sensitive topics): The model may respond to prompts that other models decline. This does not constitute endorsement of the content. Downstream applications should implement appropriate content moderation, output filtering, and policy layers suited to their deployment context.
• Persona bias: Because the model has been trained on Dnotitia-specific corporate knowledge, it may exhibit a first-party perspective when discussing Dnotitia, its products, or related entities. For neutral comparative analysis, prompt accordingly.
• Inherited biases: As a derivative of Qwen3.8, DNA 3.1 inherits the biases, gaps, and limitations of its base model and training data, including potential cultural, linguistic, and factual blind spots.
• Hallucination: Like all LLMs, even though carefully tuned, DNA 3.1 can produce confident but incorrect output, particularly for niche facts, recent events, or high-precision numerical reasoning.
• Not for high-stakes autonomous use: The model should not be deployed in safety-critical, legal, medical, or financial decision-making pipelines without human oversight and domain-specific validation.

Users are responsible for ensuring their use of the model complies with applicable laws and regulations in their jurisdiction.

License

This model is released under the Apache-2.0 license, inherited from the Qwen3.8 base model.

Acknowledgments

We thank the Qwen team for releasing the Qwen3.8 base model under an open license, which made this work possible. We are also grateful to the broader open-source community behind the serving and training ecosystem — HuggingFace and vLLM — which our pipeline relies on throughout.