AI Model Guide for Broadcasters
Not every model fits every workflow. Whether you are generating artist bios, building playlists from your media library, or writing promotional copy for your station, this guide helps you choose the right LLM for the job — matched to your hardware, your use case, and your budget.
Model Comparison Table
Side-by-side specs for every model we recommend to broadcasters. RAM and disk figures are for Q4_K_M quantized versions running through Ollama. Speed ratings assume a modern 8-core CPU with no GPU.
| Model | Parameters | RAM Required | Disk Size | Speed | Best For (Broadcasting Context) |
|---|---|---|---|---|---|
| Llama 3.1 8B | 8B | 8 GB | 4.7 GB | Medium | Artist bios, music research, show prep, general knowledge. The Swiss Army knife. |
| Llama 3.1 70B | 70B | 40 GB | 39 GB | Slow | Deep research, long-form content, complex analysis. Pro tier only. |
| Gemma 2 2B | 2B | 4 GB | 1.6 GB | Very Fast | Quick playlist logic, tag classification, lightweight tasks. Budget-friendly. |
| Gemma 2 9B | 9B | 10 GB | 5.4 GB | Medium | Playlist generation from media library context, genre analysis, scheduling suggestions. |
| Mistral 7B | 7B | 8 GB | 4.1 GB | Fast | Content writing, show descriptions, social posts, promotional copy. Excellent writer. |
| Mixtral 8x7B | 47B MoE | 26 GB | 26 GB | Medium | Multi-task: switches between research and writing seamlessly. Pro tier. |
| Phi-3 Mini | 3.8B | 4 GB | 2.3 GB | Very Fast | Analytics summaries, quick lookups, low-resource environments. |
| CodeLlama 7B | 7B | 8 GB | 3.8 GB | Fast | Plugin development, API integration scripts, automation code. |
| Qwen 2.5 7B | 7B | 8 GB | 4.4 GB | Fast | Multilingual broadcasting, international station metadata, non-English content. |
All sizes are approximate for Q4_K_M quantization. GGUF format via Ollama. GPU offloading reduces RAM needs significantly.
Best Model by Use Case
Every broadcasting task has different demands — factual accuracy, prose quality, speed, or structured-data reasoning. Here is what we recommend for the six most common AI workflows in internet radio.
Artist Bio Research
Recommended Llama 3.1 8B
Llama 3.1 8B carries broad world knowledge from its massive training corpus. It excels at factual recall — artist birthplaces, discographies, genre histories, label affiliations — and outputs coherent multi-paragraph bios suitable for on-air reading or website publishing. It handles follow-up questions well, so you can refine a bio across multiple prompts without losing context.
Playlist Generation
Recommended Gemma 2 9B
Gemma 2 9B is efficient at processing structured data like a media library database export. Feed it your track list with artist, title, genre, BPM, and play history — it recognizes patterns, avoids recent repeats, balances tempo arcs, and respects artist separation rules. Its 8K context window fits thousands of tracks as CSV, making it ideal for building hour-blocks or full-show rundowns from your actual library.
Content Writing
Recommended Mistral 7B
Mistral 7B produces polished prose with minimal editing. Station descriptions, social media posts, show promos, episode summaries, listener contest copy — it nails the tone every time. It generates fast enough for real-time workflows, so you can produce a week of social content in a single session. Pair it with a system prompt defining your station voice and it will stay on-brand consistently.
Show Prep & Scripting
Recommended Llama 3.1 8B
Talk radio hosts, morning show teams, and specialty show DJs all need prep material. Llama 3.1 8B writes conversational scripts, talking points, interview question lists, and segment outlines that sound natural when read on-air. It understands radio timing — ask for “a 90-second intro” and it delivers copy that actually fits in 90 seconds. Great for both spoken-word and music-focused show formats.
Metadata Enhancement
Recommended Phi-3 Mini
When you need to classify thousands of tracks by mood, energy level, or sub-genre, speed matters more than eloquence. Phi-3 Mini runs on as little as 4 GB of RAM and processes tag/genre classification requests at high throughput. Feed it batches of track metadata and get back normalized genre tags, mood labels, and BPM brackets. Perfect for stations cleaning up legacy libraries or ingesting new music in bulk.
Play History Analysis
Recommended Gemma 2 9B
Gemma 2 9B can ingest your play log data — timestamps, track IDs, listener counts, skip rates — and surface actionable insights. It identifies listener drop-off patterns, discovers which track sequences keep audiences engaged, and suggests programming changes based on what actually performed across any daypart or surface. Feed it a week of logs and ask “what should I change about my 6–10 AM block?” — you will get specific, data-grounded recommendations.
How to Feed Your Media Library
The real power of local AI for broadcasters is not generic chat — it is giving the model YOUR data. Export your media library and play history, then use it as context in every prompt.
The Concept
Every broadcast automation system stores your media library in a database: track title, artist, genre, BPM, duration, play count, last played date, and often mood or energy tags. Export that data as CSV or JSON, then paste it (or a filtered subset) into your prompt as context. The AI model does not need to “know” your music — you tell it what you have, and it reasons over your actual inventory.
This approach works with any model. Smaller models like Gemma 2 9B handle structured data efficiently. Larger models like Llama 3.1 8B add richer reasoning when you need explanations alongside the playlist.
Step 1: Export Your Library
From Mcaster1Studio, use the media library export feature. From other systems, run a SQL query or use the built-in export. You want at minimum: Artist, Title, Genre, Duration, BPM, Play Count, Last Played.
Step 2: Prompt with Context
Paste the export (or a filtered slice) directly into your prompt. Here is an example that builds a real playlist from your real library:
Step 3: Iterate and Refine
The model returns a structured playlist. From there you can ask follow-up questions in the same session: “Swap track 4 for something with higher energy,” “Add two more tracks to fill 8 minutes,” or “Reorder the last block so it ends on a classic.” Because the library context is already loaded, follow-ups are fast and accurate.
Pull Commands Quick Reference
Already have Ollama installed? Pull any of the recommended models with a single command. Each model downloads once and is stored locally — no API keys, no cloud dependency, no per-token costs.
Need installation help? See the Ollama Setup Guide for full instructions including hardware requirements, systemd service setup, and GPU configuration.
Ready to Get Started?
Install Ollama, pull a model, and start integrating AI into your broadcast workflow today. No cloud subscriptions. No API rate limits. Your data stays on your machine.