Minima

A strict W1.58A8 adaptation of LiquidAI/LFM2.5-Encoder-350M, built with SSHDotCodes/minima.

• Logical matrix values: {-1, 0, +1} (1.585 bits)
• Physical artifact format: I2_S, four trits per byte
• Dynamic int8 activations
• Group size 128, no recovery adapters
• Full 8,192-token encoder context
• Packed weight file: 89.9 MB (94,298,568 bytes)
• Release status: release candidate (six-task gated retention 85.45% vs FP32)

Use

pip install "minima-lfm @ git+https://github.com/SSHDotCodes/minima.git"

from minima import MinimaModel

model = MinimaModel.from_pretrained("ProCreations/minima", device="cpu")
outputs = model(input_ids=input_ids, attention_mask=attention_mask)

CPU inference defaults to a one-time FBGEMM dynamic-int8 packing of each ternary matrix. Set MINIMA_CPU_BACKEND=i2s for the direct packed 2-bit AVX2/ARM NEON kernel.

Encoder quality

The six-task downstream gate compares matched 800-step fine-tunes. Packed ternary matrices stay frozen; only the task head and non-matrix parameters adapt. Required relative mean: >= 96%.

Task	FP32	Minima	Capped retention
SST2	0.79817	0.75459	94.54%
QNLI	0.60608	0.59491	98.16%
MNLI	0.40621	0.38074	93.73%
MRPC	0.81694	0.81873	100.00%
STSB	0.54812	0.40660	74.18%
COLA	0.11382	0.05927	52.07%
Mean			85.45%

License

The weights remain subject to the LFM Open License v1.0 shipped in this repository. The Minima runtime code is MIT licensed.