WIZARDLM 30B
Evol-Instruct on LLaMA 1
Microsoft Research's Evol-Instruct fine-tune of LLaMA 1 30B. Real benchmarks: 59% MMLU, 83% HellaSwag, 62.5% ARC. A historically significant model now surpassed by smaller, newer alternatives.
Non-Commercial License Warning
WizardLM 30B is based on Meta's original LLaMA 1, which carries a non-commercial research license. You cannot use this model for commercial purposes, paid products, or revenue-generating services. For commercial use, consider Llama 3.1 8B or Mistral 7B, which have permissive licenses.
What Is WizardLM 30B
Origin and Base Model
- Team: WizardLM (Microsoft Research collaboration)
- Base Model: LLaMA 1 30B (Meta's original LLaMA, NOT LLaMA 2)
- Training Method: Evol-Instruct -- automatically evolving instruction complexity
- Release Date: June 2023
- Architecture: Decoder-only Transformer, 30B parameters
- Context Length: 2,048 tokens (hard LLaMA 1 limit)
- License: Non-commercial (LLaMA 1 restriction)
Key Specifications
Evol-Instruct: Training Methodology
Evol-Instruct is WizardLM's key innovation. Instead of relying on hand-written instruction data, the method uses an LLM to automatically rewrite simple instructions into more complex ones through evolutionary steps. This produces training data with greater depth and diversity than manual curation.
How Evol-Instruct Works
- Start with simple instructions -- basic tasks like "write a function to sort a list"
- In-depth evolving -- add constraints, increase reasoning steps, require multi-step solutions
- In-breadth evolving -- generate entirely new topics and task types
- Filter and select -- remove failed evolutions (too similar, nonsensical, or unanswerable)
- Fine-tune on evolved data -- train the base LLaMA model on the resulting complex instructions
Why It Mattered
Real Benchmark Results
MMLU Comparison (Local Models Only)
Performance Benchmarks
Multi-Benchmark Profile (Real Scores)
Performance Metrics
Benchmark Breakdown (HuggingFace Open LLM Leaderboard)
Source: Hugging Face Open LLM Leaderboard (v1, now archived), runs of TheBloke/WizardLM-30B-fp16 (the merged WizardLM-30B-V1.0 weights) dated 31 July 2023 (MMLU 5-shot, HellaSwag 10-shot, ARC 25-shot, TruthfulQA 0-shot mc2) and 22 October 2023 (WinoGrande 5-shot). The leaderboard's runs of the official WizardLM/WizardLM-30B-V1.0 repository score near chance (MMLU 23.1) because that repository holds delta weights, not a usable model.
VRAM Requirements by Quantization
Memory Usage Over Time
Quantization Options for WizardLM 30B
| Quantization | VRAM Required | Quality Loss | Best GPU | Speed ceiling* |
|---|---|---|---|---|
| Q4_K_M | ~24GB | Moderate | RTX 3090/4090 (24GB), just | ~51 tok/s (RTX 4090) |
| Q5_K_M | ~27GB | Low | RTX 5090 (32GB) / A6000 (48GB); too big for 24GB at full context | ~78 tok/s (RTX 5090) |
| Q8_0 | ~39GB | Minimal | A6000 (48GB) / M1 Ultra | ~22 tok/s (A6000) |
| FP16 | ~69GB | None | A100 (80GB) / dual A6000 | ~31 tok/s (A100 80GB) |
VRAM Required = the Ollama file (wizardlm:30b-q4_K_M 19.6 GB, q5_K_M 23.1 GB, q8_0 34.6 GB, fp16 65.1 GB; registry, 2 November 2026) + 3.3 GB of KV cache for the full 2,048-token context (60 layers × 6,656 × 2 × 2 bytes = 1.6 MB per token; LLaMA 1 has no grouped-query attention) + about 1 GB for the runtime's own buffers. Arithmetic, not a measurement. A 24GB card holds 24 GiB, about 25.8 GB in these units, so Q4_K_M fits with little to spare (less if the same card drives your display) and Q5_K_M does not.
*Ceiling = vendor memory bandwidth ÷ the file size above: RTX 4090 1,008 GB/s, RTX 5090 1,792 GB/s, RTX A6000 768 GB/s, A100 80GB SXM 2,039 GB/s. The most a dense model can generate on that card, not a measurement; real runs are slower, and much slower with any CPU offload.
Apple Silicon users: the Q4_K_M build needs about 23 GB with its cache, so a 16GB Mac cannot hold it and a 32GB Mac is tight once macOS takes its share of unified memory; 48GB or more is comfortable, and Q8_0 wants 64GB. If none of these tiers fit your machine, WizardLM 13B runs on a 12GB card at Q4 (a 7.4-7.9 GB download) and uses the same Evol-Instruct training.
System Requirements
Local Comparison: WizardLM 30B vs Alternatives
| Model | Size | RAM Required | Cost/Month |
|---|---|---|---|
| WizardLM 30B | 30B | ~24-69GB VRAM | Free (non-commercial) |
| Llama 2 13B | 13B | ~12-30GB VRAM | Free (commercial OK) |
| Vicuna 33B | 33B | ~24-69GB VRAM | Free (non-commercial) |
| Guanaco 33B | 33B | ~24-69GB VRAM | Free (non-commercial) |
| WizardVicuna 30B | 30B | ~24-69GB VRAM | Free (non-commercial) |
RAM Required = Q4_K_M to FP16 file size (Ollama registry, 2 November 2026; Guanaco 33B is not in the Ollama library, but TheBloke's GGUF files for it are the same sizes as the other LLaMA 33B models) + KV cache at each model's full context (2,048 tokens for the LLaMA 1 models, 4,096 for Llama 2 13B) + about 1 GB for runtime buffers. Arithmetic, not measured.
Context: Mid-2023 Local AI Landscape
In mid-2023, WizardLM 30B, Vicuna 33B, and Guanaco 33B were all competing for the title of best local instruction-following model. They were all LLaMA 1 fine-tunes with similar hardware requirements and non-commercial licenses. WizardLM's Evol-Instruct approach gave it an edge on complex multi-step instructions, while Vicuna excelled at conversational tasks — a merge of the two lineages shipped as WizardVicuna 30B, which appears in the table above (its modern counterparts are in our uncensored local LLM guide). Today, all of these models have been superseded by Llama 3.1 8B and Mistral 7B, which are both smaller, faster, smarter, and commercially licensed.
Installation with Ollama
Install Ollama
Download and install Ollama for your platform
Pull WizardLM 30B
Download the Q4_K_M build (19.6 GB)
Run the Model
Start a conversation with WizardLM 30B
Check VRAM Usage
Verify GPU memory allocation
Note: there is no plain wizardlm:30b tag; the Ollama library lists WizardLM 30B by quantization, for example wizardlm:30b-q4_0 (18.4 GB) and wizardlm:30b-q4_K_M (19.6 GB), checked against the Ollama registry on 2 November 2026. The GGUF files are also available on HuggingFace from TheBloke.
Example Prompt
Key Limitations to Know
2,048 Token Context
LLaMA 1's hard context limit is 2,048 tokens -- roughly 1,500 words total (input + output combined). This makes WizardLM 30B unsuitable for document analysis, long conversations, or any task requiring extended context. Modern models like Llama 3.1 offer 128K tokens.
Non-Commercial License
The LLaMA 1 base model's license prohibits commercial use entirely. You cannot use WizardLM 30B in products, paid services, or revenue-generating applications. This is a dealbreaker for most professional use cases.
High VRAM for Low Benchmark Scores
At ~24-69GB VRAM, WizardLM 30B requires significant hardware. Modern models like Llama 3.1 8B achieve higher MMLU scores (66.7 on Meta's base-model card) while requiring only 5-8GB VRAM. The performance-per-VRAM ratio is poor by 2026 standards.
Outdated Knowledge Cutoff
LLaMA 1's training data has a cutoff around early 2023. The model has no knowledge of events, technologies, or developments after that date. Combined with the short context window, this limits its practical utility.
2026 Assessment: Is WizardLM 30B Still Worth Running?
The Honest Answer: Probably Not for Production
WizardLM 30B was an important model in the history of local AI. Its Evol-Instruct methodology was genuinely innovative and influenced many subsequent projects. However, by 2026, the model has been thoroughly surpassed.
Reasons to still use it
- -- Studying the Evol-Instruct methodology
- -- Academic research comparing 2023-era models
- -- You already have it downloaded and running
- -- Nostalgic interest in early LLaMA 1 fine-tunes
Reasons to choose something else
- -- Llama 3.1 8B is faster, smarter, and commercially licensed
- -- Mistral 7B needs about 5GB VRAM vs ~24-69GB and scores higher
- -- 2048-token context is crippling for real work
- -- Non-commercial license blocks business use
- -- No active development or community support
Bottom line: WizardLM 30B is historically significant as a pioneer of instruction evolution. For any practical task in 2026, use Llama 3.1 8B, Mistral 7B, or Qwen 2.5 7B instead.
Local AI Alternatives (2026 Recommendations)
| Model | Size | MMLU | Context | VRAM (Q4) | License | Recommended? |
|---|---|---|---|---|---|---|
| WizardLM 30B | 30B | 59% | 2K | ~24GB | Non-commercial | Historical only |
| Llama 3.1 8B | 8B | 67% | 128K | ~5GB | Commercial OK | Best replacement |
| Mistral 7B | 7B | 60% | 32K | ~5GB | Apache 2.0 | Great alternative |
| Qwen 2.5 7B | 7B | 74% | 128K | ~5GB | Apache 2.0 | Highest quality |
| Phi-3 Mini | 3.8B | 69% | 128K | ~2.5GB | MIT | Best for low VRAM |
Every model above scores higher on MMLU than WizardLM 30B while using a fraction of the VRAM and offering commercial licenses. MMLU for the alternatives is each developer's published 5-shot figure (Meta's Llama 3.1 8B base card, the Mistral 7B paper, Qwen's Qwen2.5 report for the 7B base, Microsoft's Phi-3 report), so the setups differ from the leaderboard run above. VRAM (Q4) is the default 4-bit Ollama download plus a short context.
FAQ
Can I use WizardLM 30B for commercial purposes?
No. WizardLM 30B is based on LLaMA 1, which has a non-commercial research license from Meta. You cannot use it in products, paid services, or any revenue-generating application. For commercial use, switch to Llama 3.1, Mistral, or Qwen models which all have permissive licenses.
How much VRAM does WizardLM 30B need?
It depends on quantization. At the full 2,048-token context, Q4_K_M needs about 24GB (a 19.6 GB file, 3.3 GB of KV cache and runtime buffers), which just fits an RTX 3090/4090; Q5_K_M needs about 27GB, more than a 24GB card holds; Q8_0 needs about 39GB and full FP16 about 69GB. On a 24GB card, Q4_K_M is the build to use.
What is WizardLM 30B's context window?
Only 2,048 tokens. This is a hard limitation inherited from LLaMA 1. It means total input and output combined cannot exceed roughly 1,500 words. This is one of the model's biggest practical limitations. Modern models offer 32K to 128K tokens.
What is Evol-Instruct and why does it matter?
Evol-Instruct is WizardLM's training methodology where an LLM automatically evolves simple instructions into more complex ones. Starting with "write a sort function," it might evolve to "write a parallel merge sort handling edge cases with custom comparators." This was innovative in 2023 because it showed you could generate high-quality training data automatically, influencing many later models.
Is WizardLM 30B still worth downloading in 2026?
For most users, no. Models like Llama 3.1 8B, Mistral 7B, and Qwen 2.5 7B all score higher on benchmarks, use a fraction of the VRAM, have longer context windows, and come with commercial licenses. WizardLM 30B is primarily of historical and academic interest now.
What is the difference between WizardLM 30B and WizardLM 2?
WizardLM 30B (June 2023) is based on LLaMA 1 30B. WizardLM 2 (released later) used LLaMA 2 and Mistral bases with improved Evol-Instruct training, achieving significantly better results. If you want a WizardLM model, look for WizardLM 2 variants which are built on better foundations.
Build Real AI on Your Machine
RAG, agents, NLP, vision, and MLOps - chapters across 25 courses that take you from reading about AI to building AI.
Related Local Models
WizardLM 30B: Evol-Instruct Training Pipeline
WizardLM 30B training pipeline: simple instructions undergo evolutionary complexity scaling via Evol-Instruct, then fine-tune the LLaMA 1 30B base model for improved instruction following
Go from reading about AI to building with AI
25 structured courses. Hands-on projects. Runs on your machine. Start free.
Written by
Barsha Mohapatra
Founder and Editor, Local AI Master · Founder, Little AI Master
Founder of Little AI Master and the editor behind every course and guide on Local AI Master. Years of tutoring students taught her where AI education actually fails.
Continue Learning
Explore modern local AI models that have surpassed WizardLM 30B:
- PILLARLocal AI Models Directory: Every Model Compared
- DeepSeek V4 Local Setup: 1M Context, MIT (2026)
- models/llama-3-1-8b
- models/deepseek-coder-v2-236b
- Phi-3 Mini 3.8B: 69% MMLU on 8GB RAM (Microsoft SLM)
- models/gemma-2-9b
- GPT-5.6 Explained: Luna, Terra, Sol Pricing & Local Options
- StarCoder 2 15B – Technical Analysis & Performance Guide
- models/gpt-oss
- models/devstral
Grab the AI Starter Kit — career roadmap, cheat sheet, setup guide
No spam. Unsubscribe with one click.
Found your model? Now build something with it.
25 hands-on courses — RAG, agents, fine-tuning — all running locally. First chapter free, no card.