FluidInference/d1-omni-600m-coreml

d1-omni-600M · Core ML

Core ML conversion of the text path of Liquid AI's d1-omni-600M (revision 02b55d70): the LFM2.5 encoder trunk plus the decision head, answering typed questions (yes/no, choice, score) in one forward pass with zero generated tokens. The vision and audio encoders are not included.

Swift runtime and demos: FluidUse (D1OmniManager, D1OmniModelStore, ModerationDemo).

Files

file
d1-omni-text.mlpackage	one fp16 multifunction package (729 MB, weights shared)
tokenizer.json, config.json	upstream tokenizer and config (calibration temperatures)
LICENSE	LFM Open License v1.0

Functions are named L{tokens}_K{options}_B{batch}: L64/L128/L256_K2_B1 and _B8 (yes/no, one or eight questions per call) and L128/L256_K8_B1 (up to eight options). Inputs: input_ids, attention_mask (B, L), marker_pos, marker_mask (B, K), qtype (B,), all int32. Output logits (B, K): divide by the temperature in config.json and softmax over the used marker slots (a yes/no reads as [no, yes]). Prompt layout follows upstream prompt.py.

Conversion notes

• The trunk and head were rewritten with traceable ops; wrapper vs PyTorch max |Δp| 7.7e-6.
• The Neural Engine's fused silu is about 1.4% off, which compounded over 16 SwiGLU MLPs (41 of 492 Snake decisions flipped vs PyTorch). SiLU is written as x * 0.5 * (1 + tanh(x / 2)), the same function, which the Neural Engine computes accurately: 2 of 492 flips, max |Δp| 0.017.
• 99.1% of ops run on the Neural Engine (1,140 of 1,150); the rest are casts, the padding mask and the embedding lookup.

Results (Apple M5 Pro)

Toxicity on 5,000 Civil Comments test comments (CC0; clear labels: rater toxicity 0 or ≥ 0.5; prompt ≤ 256 tokens), question "Is this comment toxic?" with the Civil Comments annotators' definition of toxic, flagged at P(toxic) ≥ 0.8:

agreement with human labels	95.4% (toxic recall 67.8%, precision 82.4%)
flags that differ from PyTorch	0 of 5,000
Neural Engine only (1 per call)	102 comments/s
GPU only (8 per call)	215 comments/s
Neural Engine + GPU together	267 comments/s

Single question on the Neural Engine: about 4.6 ms at 64 tokens, 6.6 ms at 128.

License

The weights are Liquid AI's, under the LFM Open License v1.0: free to use and redistribute, but commercial use is not licensed for entities with $10M or more in annual revenue.