Sonoma-1.2B
Sonoma-1.2B is the general-availability continuation of Sonoma-1.2B Preview, derived from LiquidAI/LFM2.5-1.2B-Thinking. It was post-trained on an Apple M5 Mac with assistant-only rank-32 LoRA across all 16 model blocks, then fused into an MLX model.
Evaluation
| Checkpoint | Assistant-masked held-out loss | Perplexity |
|---|---|---|
| Sonoma-1.2B Preview | 0.934 | 2.544 |
| Sonoma-1.2B GA | 0.707 | 2.028 |
On this fixed local held-out split, the GA continuation reduces loss by 24.3% and perplexity by 20.3% versus the Preview checkpoint. This is a narrow post-training evaluation, not a general benchmark or a comparison with other models.
Deterministic behavior smoke test
The same eight-task, temperature-0 test was run with seed 20260901 and a
384-token cap on both fused MLX checkpoints:
| Checkpoint | Passed | Accuracy |
|---|---|---|
| Sonoma-1.2B Preview | 1/8 | 12.5% |
| Sonoma-1.2B GA | 3/8 | 37.5% |
This small regression suite is a directional signal only. Full per-task
responses are in eval/ga-behavior-0901.json and
eval/preview-behavior-0901.json.
Q4_K_M throughput
llama-bench build 8210 (2cd20b72e) on Apple M5, Metal, 99 GPU layers,
three repetitions, 512 prompt tokens, and 128 generated tokens measured
596.8 prompt tokens/sec and 53.0 generated tokens/sec on the Q4_K_M file.
The raw result is in eval/ga-q4-llama-bench.json; throughput varies with
context length, build, and hardware.
Training data
The reproducible selected mixture contains 7,998 deduplicated records, with assistant-token masking and a 1,024-token training limit. It draws evenly from Fable 5 Premium, OpenThoughts-114k, Glaive Code Assistant v3, and Dojo Synthetic SFT. Review each source dataset's license and terms before redistribution or use.
Files
model.safetensors: fused MLX weightsadapter/: the final rank-32 LoRA adaptergguf/Sonoma-1.2B-GA-Q4_K_M.gguf: Q4_K_M llama.cpp exporteval/ga-eval.log: independent final held-out evaluationeval/train-ga.log: training and final test logSHA256SUMS: SHA-256 hashes of the release weightstraining/: selected-data manifest and training/evaluation configssrc/andscripts/: the small reproducibility and benchmark utilities
The root weights use MLX's fused LFM2 convolution layout and are intended for
MLX. For llama.cpp, use gguf/Sonoma-1.2B-GA-Q4_K_M.gguf from this bundle or
the separate Mincofficial/Sonoma-1.2B-0901-GGUF Q4_K_M release.
License
Sonoma-1.2B is a derivative of LFM2.5-1.2B-Thinking and remains subject to the LFM Open License v1.0. Review the license before use or redistribution.