WizardLM 13B: Evol-Instruct Model
Technical review of WizardLM 13B -- the model that pioneered Evol-Instruct training. Real benchmarks from the Open LLM Leaderboard, VRAM requirements, and honest 2026 assessment.
Technical Specifications Overview
ollama run wizardlm:13bWizardLM 13B Evol-Instruct Architecture
WizardLM 13B architecture showing Evol-Instruct training pipeline: base LLaMA model fine-tuned with evolved instructions for improved instruction-following
Evol-Instruct: The Training Innovation Behind WizardLM
WizardLM's key contribution to the AI field is Evol-Instruct, a novel method for automatically generating high-complexity training data. Instead of relying on expensive human annotation (like InstructGPT) or distillation from proprietary models (like Alpaca), Evol-Instruct uses an LLM to evolve simple instructions into progressively more complex ones.
How Evol-Instruct Works
The method uses two evolution strategies described in arXiv:2304.12244:
- In-Depth Evolution: Makes instructions harder by adding constraints, increasing reasoning steps, concretizing abstract concepts, or requiring multi-step solutions
- In-Breadth Evolution: Generates entirely new instructions inspired by existing ones, expanding topic coverage and diversity
In-Depth Evolution Example
In-Breadth Evolution Example
The WizardLM team evolved 52K Alpaca instructions through multiple rounds of Evol-Instruct, producing the training dataset used to fine-tune the LLaMA base model. This approach proved effective: on the Evol-Instruct testset, WizardLM 7B showed competitive performance with ChatGPT on complex instructions while using a fraction of the parameters.
Historical note: The Evol-Instruct methodology influenced many subsequent models and training approaches. The WizardLM team later applied similar techniques to create WizardCoder and WizardMath, demonstrating the generalizability of evolved instruction tuning.
Real Benchmark Performance
All benchmark scores below are from the HuggingFace Open LLM Leaderboard -- not marketing claims.
MMLU: 13B Model Comparison
MMLU Score (%) - Source: Open LLM Leaderboard
HellaSwag & ARC Comparison
HellaSwag & ARC-Challenge (%) - Source: Open LLM Leaderboard
Multi-dimensional Benchmark Analysis
Performance Metrics
Benchmark Context
WizardLM 13B v1.2 performs comparably to other 13B-class models from the same era (mid-2023). Its MMLU of ~52% places it in the middle of the 13B pack. The model's real strength is instruction-following quality -- the Evol-Instruct training makes it handle complex multi-step instructions better than raw benchmark scores suggest. However, standard benchmarks show it does not outperform its base model (Llama 2 13B) on knowledge tasks.
Note: WizardLM was evaluated on the original Open LLM Leaderboard (v1). Scores are approximate and may vary slightly depending on evaluation framework version.
VRAM Requirements by Quantization
VRAM Usage by Quantization Level
Memory Usage Over Time
Q4_K_M
4-bit quantization, minimal quality loss
Q5_K_M
5-bit quantization, good quality balance
Q8_0
8-bit quantization, near-lossless
FP16
Full precision, no quality loss
Installation & Setup Guide
System Requirements
System Requirements
Install Ollama
Install the Ollama runtime for local model deployment
Download WizardLM 13B
Download the Q4_K_M quantized model (~7.4GB download, needs ~8GB VRAM)
Run WizardLM 13B
Start an interactive chat session with WizardLM 13B
Test instruction following
Test the model with a structured instruction-following task
Terminal Demo
Comparison with Other 13B Models
Only comparing local 13B-class models -- cloud models (ChatGPT, Claude) operate at completely different scales and are not meaningful comparisons.
| Model | Size | RAM Required | Speed | Quality | Cost/Month |
|---|---|---|---|---|---|
| WizardLM 13B v1.2 | 7.4GB (Q4) | ~8GB VRAM | ~30 tok/s | 52% | Free (Llama 2 License) |
| Llama 2 13B Chat | 7.4GB (Q4) | ~8GB VRAM | ~30 tok/s | 55% | Free (Llama 2 License) |
| Vicuna 13B v1.5 | 7.4GB (Q4) | ~8GB VRAM | ~30 tok/s | 52% | Free (Llama 2 License) |
| Nous Hermes 13B | 7.4GB (Q4) | ~8GB VRAM | ~30 tok/s | 53% | Free (Llama 2 License) |
| CodeLlama 13B | 7.4GB (Q4) | ~8GB VRAM | ~30 tok/s | 47% | Free (Llama 2 License) |
Key Takeaways
- *All 13B models share similar VRAM footprints (~8GB at Q4_K_M) since they share the LLaMA 2 architecture
- *WizardLM 13B ties with Vicuna 13B on MMLU (52%) but offers stronger instruction-following due to Evol-Instruct training
- *Llama 2 13B Chat edges ahead on raw MMLU (55%) because instruction-tuning methods can trade knowledge for alignment
- *CodeLlama 13B scores lower on MMLU (47%) because it was specialized for code, not general knowledge
Practical Use Cases
Instruction Following
WizardLM 13B excels at structured multi-step instructions thanks to Evol-Instruct training. Good for drafting documents with specific formatting, step-by-step explanations, and constrained writing tasks.
Local Privacy
All inference runs on your own hardware. No data leaves your machine. Useful for processing sensitive documents, internal communications, or proprietary information without cloud API exposure.
Learning & Experimentation
At 13B parameters, the model is small enough for experimentation. Good for learning about LLM behavior, prompt engineering practice, and understanding Evol-Instruct training effects.
What WizardLM 13B Is NOT Good For
- *Factual accuracy: With 52% MMLU, the model hallucinates frequently on knowledge-intensive tasks. Do not rely on it for factual claims without verification.
- *Production deployment: Newer models like Llama 3 8B (~66% MMLU) and Mistral 7B v0.3 (~63% MMLU) outperform WizardLM 13B while using less VRAM.
- *Code generation: Dedicated coding models like CodeLlama 13B or DeepSeek Coder 6.7B are better for programming tasks.
- *Long context: The 4,096 token context window is limiting by 2026 standards. Newer models offer 8K-128K context.
Honest 2026 Assessment
WizardLM 13B was released in July 2023 and is nearly three years old. The local AI landscape has changed dramatically since then. Here is an honest evaluation of where it stands in 2026.
Still Relevant For
- * Understanding Evol-Instruct training methodology
- * Running on older/limited hardware (8GB VRAM is accessible)
- * Historical comparison and benchmarking reference
- * Learning about instruction-tuned models
- * Offline privacy-focused use where quality is not critical
Better Alternatives in 2026
- * Llama 3 8B -- ~66% MMLU, 8K context, ~6GB VRAM
- * Mistral 7B v0.3 -- ~63% MMLU, 32K context, ~6GB VRAM
- * Phi-3 Mini 3.8B -- ~69% MMLU, 128K context, ~3GB VRAM
- * Gemma 2 9B -- ~64% MMLU, 8K context, ~7GB VRAM
- * Qwen 2.5 7B -- ~68% MMLU, 128K context, ~6GB VRAM
Bottom Line
WizardLM 13B is historically significant as the model that popularized Evol-Instruct training. For new deployments in 2026, smaller and newer models deliver substantially better performance per VRAM. If you are starting fresh, consider Llama 3 8B or Mistral 7B instead. If you are already using WizardLM 13B and it meets your needs, there is no urgent reason to switch, but upgrading will yield meaningfully better results.
WizardLM 13B Performance Analysis
Based on our proprietary 77,000 example testing dataset
Overall Accuracy
Tested across diverse real-world scenarios
Performance
~30 tokens/s on M1 MacBook Pro (Q4_K_M)
Best For
Instruction-following tasks, complex multi-step prompts, structured document generation
Dataset Insights
✅ Key Strengths
- • Excels at instruction-following tasks, complex multi-step prompts, structured document generation
- • Consistent 52%+ accuracy across test categories
- • ~30 tokens/s on M1 MacBook Pro (Q4_K_M) in real-world scenarios
- • Strong performance on domain-specific tasks
⚠️ Considerations
- • Limited context (4K tokens), outdated knowledge, lower accuracy than newer 7B-8B models
- • Performance varies with prompt complexity
- • Hardware requirements impact speed
- • Best results with proper fine-tuning
🔬 Testing Methodology
Our proprietary dataset includes coding challenges, creative writing prompts, data analysis tasks, Q&A scenarios, and technical documentation across 15 different categories. All tests run on standardized hardware configurations to ensure fair comparisons.
Want the complete dataset analysis report?
Was this helpful?
Frequently Asked Questions
What are WizardLM 13B's actual benchmark scores?
WizardLM 13B v1.2 scores approximately 52% on MMLU, 79% on HellaSwag, and 57% on ARC-Challenge according to the HuggingFace Open LLM Leaderboard. These are real evaluation scores, not marketing claims. It is competitive with Llama 2 13B Chat (55% MMLU) and Vicuna 13B v1.5 (52% MMLU) in the 13B parameter class.
How much VRAM does WizardLM 13B need?
WizardLM 13B needs approximately 8 GB VRAM at Q4_K_M quantization, 10 GB at Q5_K_M, 14 GB at Q8_0, and 26 GB at full FP16 precision. A 16GB GPU (RTX 4060 Ti 16GB) or Apple M1/M2 with 16GB unified memory can run Q4_K_M or Q5_K_M comfortably. Install via Ollama: ollama run wizardlm:13b
What is Evol-Instruct and how does it train WizardLM?
Evol-Instruct is a novel training methodology introduced in the WizardLM paper (arXiv:2304.12244). It evolves simple instructions into increasingly complex ones through two strategies: in-depth evolution (adding constraints, deepening reasoning, concretizing) and in-breadth evolution (generating new related topics). This creates diverse, high-complexity training data without expensive human annotation.
Is WizardLM 13B still worth using in 2026?
WizardLM 13B is a 2023 model that has been surpassed by newer alternatives. For instruction-following, consider Llama 3 8B (~66% MMLU, 8GB VRAM) or Mistral 7B v0.3 (~63% MMLU, 6GB VRAM), both of which outperform WizardLM 13B while using less VRAM. WizardLM 13B remains historically significant for pioneering the Evol-Instruct methodology.
What license does WizardLM 13B use?
WizardLM 13B v1.2 uses the Llama 2 Community License from Meta, which permits commercial use provided you comply with Meta's acceptable use policy and have fewer than 700 million monthly active users. Earlier versions (v1.0, v1.1) used the original LLaMA license which was more restrictive.
Resources & References
Official Sources
Research Papers
Benchmarks & Tools
Written by Pattanaik Ramswarup
AI Engineer & Dataset Architect | Creator of the 77,000 Training Dataset
I've personally trained over 50 AI models from scratch and spent 2,000+ hours optimizing local AI deployments. My 77K dataset project revolutionized how businesses approach AI training. Every guide on this site is based on real hands-on experience, not theory. I test everything on my own hardware before writing about it.
Related Guides
Continue your local AI journey with these comprehensive guides