Downloads · 30 days
16
4% of all-time downloads
LFM2.5-8B-A1B-MLX-nvfp4
bkideas/LFM2.5-8B-A1B-MLX-nvfp4
LFM2.5-8B-A1B-MLX-nvfp4 is a machine learning model from bkideas. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for mlx. The card lists the license as other.
This model is a quantized NVFP4 MLX variant of LiquidAI/LFM2.5‑8B‑A1B‑MLX‑bf16, created by LiquidAI. Original model licensed under the LiquidAI Model License.
All-time downloads
380
Public
Parameters
8.5B
4.8 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors4.8 GB · 100%
Parameter types
How the weights are stored.
U328.5B · 100%
Base models
Model card
From the Hugging Face model README
This model is a quantized NVFP4 MLX variant of LiquidAI/LFM2.5‑8B‑A1B‑MLX‑bf16, created by LiquidAI. Original model licensed under the LiquidAI Model License.
NVFP4 MLX Quantization — Performance & Quality
This model is a 4‑bit NVFP4 MLX‑quantized variant of the original BF16 LFM2.5‑8B‑A1B model. NVFP4 is MLX’s optimized 4‑bit format designed for efficient inference on Apple Silicon GPUs.
Why NVFP4?
NVFP4 reduces memory usage by ~65% and increases generation speed by ~1.6–1.8× on M‑series chips, while preserving most of the model’s quality.
Performance Comparison (Representative MLX Benchmarks)
| Metric | BF16 | NVFP4 | Notes |
|---|---|---|---|
| Memory usage | ~15 GB | ~5 GB | Fits on 16 GB Macs |
| Token speed (M5 Max) | ~41 tok/s | ~72 tok/s | ~1.75× faster |
| Perplexity | 1.00× | 1.02–1.03× | ~2–3% degradation |
| Output quality | Baseline | ~95–98% identical | Minor reasoning loss |
Pros
- Much lower memory footprint
- Faster inference on macOS
- Lower power usage
- Ideal for laptops and smaller RAM configs
Cons
- Slight quality degradation (1–3%)
- Not suitable for fine‑tuning
- Slightly more drift in very long generations
Practical Impact
For chat, summarization, and coding, NVFP4 behaves almost identically to the BF16 model.
For math/logic‑heavy tasks, BF16 remains slightly more accurate.
- Library
- mlx
- Type
- lfm2_moe
- License
- other
- Created
- Jun 18, 2026
- Updated
- Jun 21, 2026