Model Catalog Reference
Catalog schema, model family status, and compatibility guidance
Halo Forge has two different model concepts:
- Model catalog: curated upstream/base models you can train, evaluate, serve, or use as starting points.
- Run bundles: saved groups of your trained runs in the dashboard, used for comparison and cohort eval.
Use the catalog from CLI, API, or dashboard:
halo-forge models list
halo-forge models list --mode raft --backend rocm
halo-forge models show LiquidAI/LFM2.5-350M
curl http://127.0.0.1:8000/api/public/models | jq
curl "http://127.0.0.1:8000/api/public/models?mode=vlm&status=experimental" | jq
Catalog Fields
Each entry includes:
| Field | Meaning |
|---|---|
id | Hugging Face, MLX, or local-compatible model identifier |
provider, family | Human grouping for browsing and filtering |
modalities | text, code, vision, audio |
tasks | Practical uses such as raft, preference, asr, agentic |
trainer_support | Halo Forge trainers expected to work |
backend_support | cuda, rocm, mps, mlx, cpu compatibility hints |
memory_tier | Tiny, small, medium, or large |
status | recommended, compatible, experimental, or deprecated |
known_caveats | Warnings that should affect operator choice |
mlx_variant | MLX-format sibling when known |
Status Definitions
| Status | Use it when |
|---|---|
recommended | It is a good default for at least one Halo Forge workflow |
compatible | It should work, but is not the first pick |
experimental | Interesting but not fully proven in Halo Forge |
deprecated | Kept for migration context only |
Family Guidance
Qwen and Qwen Coder
Qwen is the safest first choice. Use Qwen Coder for code SFT/RAFT, Qwen Instruct for preference tuning, and Qwen-VL for current VLM workflows.
Llama, Mistral, Gemma
Good general-purpose baselines. Watch license gates and chat-template differences.
DeepSeek, StarCoder2, CodeLlama
Useful code alternatives. Test a short verifier-backed run before long jobs because tokenizer and dependency behavior can differ by release.
Whisper
Current audio training defaults are Whisper-compatible. Start with openai/whisper-tiny for smoke tests and openai/whisper-small for useful ASR adaptation.
MLX Community
Use MLX-format repos on Apple Silicon when selecting the MLX backend. Quantization is baked into the artifact; it is not bitsandbytes runtime quantization.
Liquid AI LFM
Liquid LFM2.5 models are small, edge-oriented, and interesting for structured output, tool use, extraction, reasoning, and on-device deployment. Halo Forge lists them with experimental status until each path is tested end-to-end.
Recommended first Liquid experiments:
| Model | Use | Caveat |
|---|---|---|
LiquidAI/LFM2.5-350M | Tiny structured output and tool-use experiments | Liquid’s model card says it is not recommended for knowledge-heavy tasks or programming |
LiquidAI/LFM2.5-1.2B-Instruct | Small chat/tool-use experiments | Verify trainer behavior before long runs |
LiquidAI/LFM2.5-1.2B-Thinking | Reasoning experiments | Treat as experimental until benchmarked locally |
LiquidAI/LFM2.5-VL-450M / LFM2.5-VL-1.6B | Visual extraction experiments | Halo Forge VLM adapters need validation |
LiquidAI/LFM2.5-Audio-1.5B | ASR/TTS experiments | Halo Forge audio path is currently Whisper-oriented |
Primary references:
- Liquid model overview
- LFM2.5 announcement
- Liquid text models
- Liquid audio models
- LiquidAI/LFM2.5-350M on Hugging Face
Compatibility Is A Starting Point
The catalog answers “what should I try first?” It does not replace a smoke run. Before committing hours to a model:
halo-forge test --level smoke
halo-forge models show MODEL_ID
halo-forge sft train --model MODEL_ID --dataset codealpaca --epochs 1 --max-samples 50 --output models/smoke
Then evaluate the artifact before scaling samples, model size, or cycles.