Granite-4.0-H-Tiny — MLX 6-bit (Apple Silicon)

Maintainer / Publisher: Susant Achary

This repository provides an Apple-Silicon MLX build of IBM Granite-4.0-H-Tiny quantized to 6-bit.
Among MLX quant variants, 6-bit offers the highest fidelity while still fitting comfortably on modern M-series Macs. If your workload involves precise extraction, structured outputs, or long contexts, 6-bit is usually the best on-device choice.

🔢 Choosing a quantization level (LMX variants)

Use this table as a practical guide for a ~7B hybrid MoE LM on Apple Silicon. (Figures vary by device/context.)

Variant	Typical Peak RAM	Relative Speed	Typical Behavior	When to Choose
2-bit	~3–4 GB	🔥🔥🔥🔥	Smallest, most lossy	Minimal RAM devices; smoke tests
3-bit	~5–6 GB	🔥🔥🔥🔥	Direct, concise	Great default on M1/M2/M3/M4
4-bit	~6–7.5 GB	🔥🔥🔥	Better detail retention	If 3-bit misses details
5-bit	~8–9 GB	🔥🔥☆	Higher fidelity	Heavier docs/structured outputs
6-bit (this repo)	~9.5–11 GB	🔥🔥	Highest MLX fidelity	Best quality on-device if RAM permits

Tips

Prefer 6-bit when you have ~10–12 GB free and want maximum quality.
Use 3-bit/4-bit for tighter RAM with good latency and strong baseline quality.
For JSON/structured extraction, consider temperature 0.0 and schema-style prompts.

🔎 About Granite 4.0 (context for this build)

Architecture: Hybrid Mamba-2 + softmax attention; H tiers add Mixture-of-Experts (MoE) for sparse activation and efficiency.
Model tier: H-Tiny (~7B total params with ~1B active via MoE) — designed for long-context use and efficient serving.
License: Apache-2.0 (permissive, enterprise-friendly).
Use cases: Instruction following, long-context assistants, RAG backends, structured outputs.

This card documents the MLX 6-bit conversion. For lower-RAM devices, see the 2/3/4/5-bit guidance below.

📦 Contents of this repository

config.json (MLX), mlx_model*.safetensors (6-bit shards)
Tokenizer files: tokenizer.json, tokenizer_config.json
Any auxiliary metadata (e.g., model_index.json)

This build targets macOS on Apple Silicon (M-series) using Metal/MPS.

✅ Intended use

High-fidelity instruction following and summarization
Long-context reasoning and retrieval-augmented generation (RAG)
Structured extraction (JSON, key–value) and document parsing
On-device prototyping where answer faithfulness matters

⚠️ Limitations

As with any quantization, small regressions vs FP16 can occur (complex math/code).
Token limits and KV-cache growth still apply for very long contexts.
Always add your own guardrails/safety for sensitive deployments.

🚀 Quickstart (CLI — MLX)

Deterministic generation

python -m mlx_lm.generate \
  --model <this-repo-id> \
  --prompt "Summarize the following meeting notes in 5 bullet points:\n<your text>" \
  --max-tokens 256 \
  --temperature 0.0 \
  --device mps \
  --seed 0