mlx-community/LFM2-8B-A1B-6bit-MLX

LFM2-8B-A1B — MLX 6-bit (Apple Silicon)

Maintainer / Publisher: Susant Achary
Upstream model: LiquidAI/LFM2-8B-A1B
This repo (MLX 6-bit): mlx-community/LFM2-8B-A1B-6bit-MLX

This repository provides an Apple-Silicon-optimized MLX build of LFM2-8B-A1B at 6-bit quantization.
Among quantized tiers, 6-bit is a strong fidelity sweet-spot for many Macs—noticeably smaller than FP16/8-bit while preserving answer quality for instruction following, summarization, and structured extraction.

---

🔎 What is LFM2-8B-A1B?

• Architecture: Mixture-of-Experts (MoE) Transformer.
• Size: 8B total parameters with ~1B active per token (A1B ≈ “1B active”).
• Why MoE? At each token, a subset of experts is activated, reducing compute per token while keeping a larger parameter pool for expressivity.

Single-device memory reality: Even though only ~1B are active per token, all experts typically reside in memory during inference on one device. That means RAM planning should track total parameters, not just the active slice.

---

📦 What’s in this MLX build

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

Target: macOS on Apple Silicon (M-series) with Metal/MPS.

---

✅ Intended use

• General instruction following, chat, and summarization
• RAG and long-context assistants on device
• Schema-guided structured outputs (JSON)

⚠️ Limitations

• Quantization can cause small regressions vs FP16 on tricky math/code or tight formatting.
• For very long contexts and/or batching, the KV-cache can dominate memory—tune max_tokens and batch size.
• Add your own safety/guardrails for sensitive deployments.

---

🔢 RAM planning (6-bit, MoE, MLX)

You asked to assume and decide realistic ranges. The following are practical starting points for a single-device MLX run; validate on your hardware.

Rule-of-thumb components

• Weights (6-bit): ≈ total_params × 0.75 byte → for 8B params ≈ ~6.0 GB