bineric / lynx-instruct-30b

Bineric Lynx Instruct 30B

A European large language model with exceptional Nordic language performance.

About Bineric

Bineric is an AI company based in Oslo, Norway, built from a European perspective. We started Bineric to make AI usable for organizations that care about governance, language, and where their systems and data actually live.

Lynx is our flagship model — designed to serve European users with strong multilingual support and exceptional Nordic language performance.

Overview

Lynx is built on Qwen3-30B's efficient Mixture-of-Experts architecture. It retains strong multilingual capabilities across 100+ languages including all major European languages , while being specifically fine-tuned and rigorously evaluated on Nordic languages (Norwegian, Swedish, Danish, Icelandic) where it demonstrates exceptional results.

Key features:

• Strong European language support inherited from Qwen3 base model
• Fine-tuned and optimized for Nordic language understanding and generation
• Efficient MoE architecture: only 3B parameters active per token
• Available in 8-bit and 4-bit quantized variants for flexible deployment
• 262K context window for long-document processing

Try Lynx

Lynx is available through multiple channels:

Evaluation Results

Evaluated using EuroEval benchmark framework (March 2026).

Note: While Lynx supports all European languages via its Qwen3 base, we have rigorously evaluated performance on Nordic languages. Benchmarks for additional European languages coming soon.

Nordic Language Performance

Task Performance by Language

Norwegian (8-bit)

Danish (8-bit)

Swedish (8-bit)

Icelandic (8-bit)

Quantization Comparison (Norwegian)

8-bit quantization consistently outperforms 4-bit by ~2% on average.

Strengths & Limitations

Strengths

• Named Entity Recognition : Consistently strong across all languages (63-71% F1)
• Reading Comprehension : Excellent for Danish and Swedish (72%+)
• Knowledge Tasks : Outstanding on Danish Citizen Tests (79.3%)
• Summarization : Stable 63-66% BERTScore across all languages

Limitations

• Linguistic Acceptability : Grammatical judgment tasks are weak (10-36% MCC)
• Icelandic Common Sense : Winogrande-is performance is low (9.7%)
• Norwegian Idioms : Room for improvement (17-19% MCC)

Quantization Options

Usage

Basic Inference

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "bineric/lynx-instruct-30b",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("bineric/lynx-instruct-30b")

messages = [
    {"role": "user", "content": "Hva er hovedstaden i Norge?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With Thinking Mode

Lynx supports extended thinking for complex reasoning tasks:

messages = [
    {"role": "user", "content": "Forklar forskjellen mellom bokmål og nynorsk."}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True  # Enable reasoning mode
)

vLLM Deployment

vllm serve bineric/lynx-instruct-30b \
    --tensor-parallel-size 1 \
    --max-model-len 32768 \
    --quantization awq  # For 4-bit

Model Architecture

Qwen3 MoE Architecture
├── Total Parameters: 30.5B
├── Active Parameters: ~3B per token
├── Hidden Layers: 48
├── Hidden Size: 2048
├── Attention Heads: 32
├── KV Heads: 4 (Grouped Query Attention)
├── Experts: 128 total
├── Active Experts: 8 per token
├── Vocab Size: 151,936
└── Context Length: 262,144 tokens

Training

Lynx is fine-tuned from Qwen3-30B-A3B-Instruct with additional training on Nordic language data to improve performance on Norwegian, Swedish, Danish, and Icelandic tasks.

Citation

@misc{bineric2026lynx,
  title={Bineric Lynx: A European Large Language Model},
  author={Bineric AI},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/bineric/lynx-instruct-30b}
}

Links

• Base Model: Qwen3-30B-A3B-Instruct

---

Built with care in Oslo by Bineric