LLM Foundry

LLM Foundry is the workshop around an LLM.

It is not the model itself. It is the memory, compression, tokenizer, tool use, traces, harnesses, training plumbing, and API orchestration that make a model useful.

What it gives you

• a backend abstraction for many model sources
• tokenizer support

Provider options

• qwen for the built-in Hugging Face Qwen path
• openai for OpenAI-compatible BYOK providers
• anthropic for Anthropic BYOK providers
• hf for any Hugging Face local model
• multi or openai-multi for failover bundles

Example:

python -m llm_foundry demo --provider qwen --model Qwen/Qwen2.5-0.5B-Instruct --prompt "Hello"

Proof mode

Use proof for a terminal-first demo that shows the question, tool call trace, final answer, and JSON artefact. It is designed for screenshots and file explorer use on Windows, macOS, and Linux.

Example:

python -m llm_foundry proof --provider qwen --model Qwen/Qwen2.5-0.5B-Instruct

Proof images

• Prompt + answer
• Agent trace
• Tests
• Path handling
• Diff
• Terminal

GitHub Pages

A live GitHub Pages demo can host the docs site from docs/. Use it to try the default Qwen model path or BYOK providers.

• Default model: qwen
• BYOK providers: openai, anthropic
• Multi-endpoint bundles: multi / openai-multi

For the live page, use the commands on docs/index.html.