d1-3B Core AI ANE
Text-only LiquidAI/d1-3B decision model conversion for inference on Apple’s Neural Engine via Core AI, used by the tinkerbox runtime. Tokenization, embeddings and option readout run on the CPU, the backbone runs on Apple's Neural Engine.
- Platform: macOS 27+, Apple Silicon (M4–M6, all variations)
- Chunk size: 128 tokens. Context limit: 32,768 rendered tokens.
- Nine cache capacities share one set of weights.
- Upstream checkpoint:
LiquidAI/d1-3B@051bcc464b01b9f92942b364d9586b0ef5912432. - Source hash:
2b64ab9d9d9a03dfa3476ab6ae33fa248eee97c9710e6f16f725e25db77e8629.
Tested metrics
Apple M5 Pro (24 GB), macOS 27.0.1; measured October 16, 2026.
| Metric | Result |
|---|---|
| Top-answer agreement with official FP32 model | 100% (204/204) |
| Absolute probability drift vs FP32, mean / maximum | 0.000655 / 0.012448 |
| Single-question latency, 26–251 rendered input tokens, median / p95 | 38.9 ms / 79.0 ms |
| 4,096-token latency, median | 1.56 s |
| 32,768-token latency | 48.7 s |
From the tinkerbox-ane source checkout:
from tinkerbox_ane import d1
with d1.load() as model:
result = model.system_one(
"I was charged twice this month.",
{"refund": {"type": "noul", "instructions": "Is this a refund request?"}},
)
print(result)
Weights and converted assets use the LFM Open License v1.0. See NOTICE.