Downloads · 30 days
11
10% of all-time downloads
SmolLM3-3B-a1lamp-4bit
ageyko/SmolLM3-3B-a1lamp-4bit
SmolLM3-3B-a1lamp-4bit is a text generation model from ageyko. Use it when you need the model to write or continue text. It is set up for mlx. The card lists the license as apache-2.0.
4-bit MLX quantization of SmolLM3-3B with the A1-lamp Task-LoRA fused into the base weights. Research artifact for an on-device LLM personalization study (two-stage LoRA: a shared Task-LoRA + a per-user User-LoRA trai…
All-time downloads
106
Public
Parameters
3.1B
1.7 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors1.7 GB · 99%
Parameter types
How the weights are stored.
U323.1B · 100%
Try a prompt
Base models
Model card
From the Hugging Face model README
SmolLM3-3B-a1lamp-4bit
4-bit MLX quantization of SmolLM3-3B with the A1-lamp Task-LoRA fused into the base weights. Research artifact for an on-device LLM personalization study (two-stage LoRA: a shared Task-LoRA + a per-user User-LoRA trained on-device).
- Base:
HuggingFaceTB/SmolLM3-3B(bf16, frozen). - Fused adapter: A1-lamp Task-LoRA (r=4, α=8, all 7 projections), 1-epoch
checkpoint-1000, trained on LaMP-{3,4,7} with a BM25 top-4 profile in the
systemslot. Merged into the base with PEFTmerge_and_unload, then quantized to 4-bit (4.5 bits/weight) viamlx_lm convert. - Intended use: the frozen deployment base onto which a fresh per-user User-LoRA (r=8, q+v) is trained on device. Personal User-LoRAs are trained on-device and are not part of this repo.
Thinking mode is off in the study's regime (SmolLM3 emits <think>…</think>;
disable via the chat template's /no_think).
Load with MLX:
from mlx_lm import load, generate
model, tokenizer = load("ageyko/SmolLM3-3B-a1lamp-4bit")
- Task
- Text Generation
- Library
- mlx
- Type
- smollm3
- License
- apache-2.0
- Languages
- en
- Created
- Jul 6, 2026
- Updated
- Jul 6, 2026