Minima

A 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 32 with rank-128 FP16 recovery adapters
• Full 8,192-token encoder context
• 223.9 MB weight file, 84.2% smaller than the 1,418.0 MB source weight file

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 effective ternary-plus-recovery matrix, after which source projection tensors are released. Set MINIMA_CPU_BACKEND=i2s for the direct packed 2-bit AVX2/ARM NEON kernel. The direct I2_S path has the smaller projection representation; the FBGEMM path is the measured throughput default.

Measured CPU results

Hugging Face cpu-performance, Linux x86-64, FBGEMM, 16 threads, one warmup and five measured runs:

Sequence	FP32 median	Minima median	Speedup	Peak RSS reduction
128	181.62 ms	80.82 ms	2.25x	23.81%
512	479.12 ms	247.94 ms	1.93x	26.77%
2,048	1,402.74 ms	1,280.92 ms	1.10x	26.94%
8,192	7,878.43 ms	7,312.03 ms	1.08x	24.42%

Peak RSS includes framework and activation memory, so it does not shrink by the same 84.2% as the weight file. Raw reports are in the results dataset.

Encoder quality

The six-task downstream gate retained 96.66% of the matched FP32 baseline after per-task ratios were capped at 100%. Five non-CoLA tasks averaged 98.05%; CoLA retained 89.70%. This misses the declared 97% threshold by 0.34 percentage points, so this artifact is a release candidate, not a quality-gated release. The validation-selected CoLA schedule and all eight candidates are published in the raw report.

CUDA status

The fused Triton path avoids materializing a full dequantized weight tensor and passes correctness checks, but the current kernel is slower than upstream BF16 on an H200. It is included for optimization work, not advertised as a GPU speedup.

Diagnostics

The release distillation probe measured hidden-state cosine 0.8491, relative L2 0.5877, and 99.67% masked-token top-1 agreement across 601 positions. These are diagnostics; the downstream task gate above is the release-quality measure.

License

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