v0.0.2 — GGUF all versions (tiny-50M + base-1B)

v0.0.3 — GGUF weights, sob version

tiny-50M (56M params, ctx 2048)

• tiny-50M-q8.gguf (59.9 MB) — best quality
• tiny-50M-q4.gguf (31.9 MB) — smallest, cholbe low RAM e

base-1B (1.1B params, ctx 2048)

• base-1B-q8.gguf (1.19 GB) — best quality
• base-1B-q4.gguf (632 MB) — recommended, speed + size balance

Run

llama-server -m base-1B-q4.gguf -c 2048
llama-server -m tiny-50M-q8.gguf -c 2048

Note

• Tokenizer: 32k BPE (alfa-32k.json, v0.0.2 release e ache)
• code-100k (1.39B, Kaggle-trained) GGUF ashle next release e add hobe — HF upload pending
• Source weights: https://huggingface.co/mrsandip/Alfa-Code

Tiny-50M v0.0.1 (preview)

Preview/demo weights for pipeline + UI testing (50M Llama, BPE, 30 demo rows, CPU smoke-train). Not quality weights.

Full Colab-trained weights (1500 rows, GPU): Hub mrsandip/Alfa-Code outputs/tiny-50M.

Run: llama-server -m tiny-50M-q8.gguf -c 2048 (Q4 ~32MB for small VRAM).

Next: 1B model (v0.0.2).