Model catalog (validated)

A running list of models validated with heim — copy-paste the heim library pull commands to rebuild a library on a new drive or laptop. Everything the library holds is just files under HEIM_LIBRARY_ROOT, so "rebuild" = set that root, then run these pulls. This page grows as models are validated; if a model isn't here, it hasn't been confirmed working (not that it can't).

Rebuild in one shot

Set your library, then paste a block below:

export HEIM_LIBRARY_ROOT=/Volumes/LLM/Library   # your drive
# then paste the chat + voice pull commands you want

Chat / LLM models

Pulled from the Hugging Face Hub (heim picks a balanced GGUF quant, or the MLX build for mlx-community repos). MLX builds are Apple-Silicon-only; GGUF runs everywhere via llama.cpp.

# --- small (fast, run alongside a voice model) ---
heim library pull huggingface unsloth/Qwen3.5-4B-GGUF               # 2.6 GB · tool-capable
heim library pull huggingface lmstudio-community/Qwen3.5-4B-MLX-4bit # 2.9 GB · MLX
heim library pull huggingface unsloth/Llama-3.2-3B-Instruct-GGUF    # ~2 GB · tiny starter

# --- mid (capable all-rounders) ---
heim library pull huggingface lmstudio-community/gemma-4-12B-it-QAT-GGUF  # 6.7 GB · tools+vision+audio
heim library pull huggingface deepreinforce-ai/Ornith-1.0-9B-GGUF        # 5.2 GB
heim library pull huggingface TheDrummer/Rocinante-X-12B-v1-GGUF         # 7.0 GB · roleplay
heim library pull huggingface unsloth/gpt-oss-20b-GGUF                   # 10.8 GB · strong reasoning + tools

# --- large (big-context / coding / vision) ---
heim library pull huggingface unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF  # 17.3 GB · coding
heim library pull huggingface lmstudio-community/Qwen3.6-27B-GGUF        # 16.3 GB
heim library pull huggingface mlx-community/Qwen3.6-27B-4bit             # 15.0 GB · MLX
heim library pull huggingface lmstudio-community/gemma-4-31B-it-QAT-GGUF # 17.6 GB · vision
heim library pull huggingface lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit # 14.6 GB · MLX vision
heim library pull huggingface TheDrummer/Cydonia-24B-v4.3-GGUF          # 13.3 GB
heim library pull huggingface mradermacher/MN-Violet-Lotus-12B-GGUF     # 12.1 GB

Voice models

heim library pull voice <id> (downloads, or fast-copies from another reachable library). TTS speaks; STT transcribes. Except kokoro (ONNX, cross-platform) and outetts (llama.cpp), these run on mlx-audio (Apple Silicon).

# --- TTS ---
heim library pull voice kokoro       # Kokoro-82M · 54 named voices · the in-chat voice · cross-platform
heim library pull voice soprano      # 80M · tiny high-quality English (Kokoro-family)
heim library pull voice chatterbox   # expressive · emotion/exaggeration control + cloning
heim library pull voice spark        # 0.5B · English + Chinese
heim library pull voice csm          # 1B · voice cloning from a reference clip
heim library pull voice dia          # 1.6B · nonverbal cues + cloning + multi-speaker (seed it)
heim library pull voice outetts      # 500M · GGUF via llama.cpp · cross-platform

# --- STT (speech-to-text) ---
heim library pull voice whisper        # large-v3-turbo · multilingual · the default
heim library pull voice parakeet       # 0.6B · fast · English + 25 EU languages
heim library pull voice qwen3-asr      # 1.7B · multilingual
heim library pull voice distil-whisper # faster distilled Whisper (English)

Validated (2026-07-04) — voice detail

id	kind	mode	backend	notes
kokoro	TTS	preset	kokoro-onnx	54 voices, 8 languages; the in-chat default; cross-platform
soprano	TTS	preset	mlx-audio	80M, tiny English, Kokoro-family
chatterbox	TTS	preset	mlx-audio	emotion/exaggeration control + cloning
spark	TTS	preset	mlx-audio	English + Chinese (needs soxr, in the voice extra)
csm	TTS	clone	mlx-audio	cloning — pass a reference clip + transcript
dia	TTS	seeded	mlx-audio	expressive, nonverbal cues, multi-speaker; pin a seed
outetts	TTS	preset	llama-tts	GGUF; cross-platform
whisper	STT	—	mlx-audio	multilingual default
parakeet	STT	—	mlx-audio	fast; EN + 25 EU languages
qwen3-asr	STT	—	mlx-audio	multilingual
distil-whisper	STT	—	mlx-audio	faster English Whisper

Not yet working (load but produce no audio via mlx-audio's high-level API, or need bespoke args): qwen3-tts, KittenTTS, OuteTTS-1.0 (mlx), Qwen3-TTS-VoiceDesign, Voxtral-TTS. Tracked in the roadmap (ROADMAP.md in the repo); revisit as mlx-audio adds support.

Benchmarks

Tool-calling and code/reasoning are measured by the heim benchmark suites (tools-datetime, tools-utils, tools-search, tools-web, tools-dhis2, python, rust). Run heim benchmark to score a model against them; use the results to pick which chat model to bind in a project.

DHIS2 (tools-dhis2, read-only): the standout is deepreinforce-ai/Ornith-1.0-9B-GGUF — a perfect 12/12 driving the DHIS2 CLI bridge, and the fastest (~12s/problem) and smallest (5.2 GB) model to do so, beating the 27B/31B models. gemma-4-12B and Qwen3.6-27B also went 12/12; Qwen3.5-4B (2.6 GB) and gpt-oss-20b reached 11/12. Roleplay finetunes (Rocinante, MN-Violet-Lotus) scored 0/12 — they never call the tool. See the full write-up in DHIS2 benchmark report.

DHIS2 (tools-dhis2-write, read-write): driving writes (create / rename / delete against a local instance) is far harder, and bigger does not help. The small gemma-4-12B-it-QAT leads at 4/7; Qwen3-Coder-30B and Qwen3.6-27B tie at 3/7, gpt-oss-20b 2/7, and both Ornith-1.0-9B and the biggest model, the 35B-A3B MoE Qwen3.6-35B-A3B, trail at 1/7 (the 35B also left the most residue). The multi-step create→delete→confirm lifecycle trips up every model — the largest ones over-generate and drop the completion protocol — so no local model is yet trustworthy for unattended DHIS2 writes.