Sonoma-1.2B-Preview

Sonoma-1.2B Preview

Sonoma-1.2B Preview is an early compact reasoning and instruction model 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. This repository includes fused MLX weights, the selected LoRA adapter, an F16 GGUF, and a 697 MiB Q4_K_M GGUF.

The release improves on its untouched base in the local held-out evaluation used during training. That result is useful, but narrow. It does not establish that Sonoma beats larger models or frontier systems.

Results

The held-out split has 77 examples. Sonoma reduced loss by 22.3% and perplexity by 18.9% relative to the base. The eight-task test is a deterministic regression check, not a general benchmark.

Read the system card for the training mixture, selection decision, evaluation details, known limitations, and safety notes.

Use with MLX

pip install mlx-lm
mlx_lm.generate \
  --model Mincofficial/Sonoma-1.2B-Preview \
  --prompt "Solve 5x - 7 = 53."

Use the Q4_K_M GGUF

hf download Mincofficial/Sonoma-1.2B-Preview \
  gguf/Sonoma-1.2B-Q4_K_M.gguf \
  --local-dir .

llama-cli \
  -m gguf/Sonoma-1.2B-Q4_K_M.gguf \
  -cnv -ngl 99 -c 4096

The root weights use MLX's fused LFM2 convolution layout and are intended for MLX. Use the GGUF files for llama.cpp. They are not advertised as a drop-in Transformers checkpoint.

Files

• model.safetensors: fused MLX model
• gguf/Sonoma-1.2B-Q4_K_M.gguf: quantized llama.cpp release
• gguf/Sonoma-1.2B-F16.gguf: full-precision GGUF
• adapter/: selected rank-32 LoRA at release scale 8
• eval/: metrics and deterministic smoke-test reports
• SHA256SUMS: hashes for the four weight artifacts

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.