D1 native checkpoint qualification — 2026-10-09

Both pinned, original LiquidAI safetensors ran through the native Swift/MLX text decide CLI without conversion. D1-3B was tested in BF16 with text and images; D1 omni-600M was tested in FP32 with text, images, and audio. Each of five requests contained one yes/no, one choice, and one ordinal score question: 15 decisions total. The runtime computation is from commit b7a4bde5.

Original weight identity

Downloads matched the SHA-256 hashes published by Hugging Face for these exact revisions. No converted weights, quantization, or Python inference fallback were used by the native runtime.

Model	Revision	Weight bytes	SHA-256
D1-3B	da1fe36a861f24690f27f622dca1d8688503d113	6,247,065,504	50e03317847caf6df9a9aee27ed40f20554a86a21e60d1d47ba41a422b546c0c
D1 omni-600M	414f8d6438174f5b2133a9c21a478fc42625e308	2,348,774,500	0713bb05270c2685ad106522f4092bceeeb3a93cf79b401f399a712296c911e1

Comparison results

The independent reference loads the original checkpoint's Python implementation and weights on CPU, using BF16 for D1-3B and FP32 for omni. It executes each question separately to match native execution. Reference generation is test infrastructure; user inference requires no Python installation.

Model	Input	Maximum absolute probability difference
D1-3B BF16	Text	0.001278479
D1-3B BF16	Image	0.000642697
Omni FP32	Text	0.000001252
Omni FP32	Image	0.000008881
Omni FP32	Audio	0.000002027

Choice identities, yes/no classifications at 0.5, and executed input-token counts agree for all cases. All CLI commands exited successfully and emitted valid probability distributions with zero output tokens. Expected score values agree within the committed numerical tolerances. Native outputs and comparison receipt retain every answer.

Inputs are in Tests/MereRunCoreTests/Fixtures/D1/released/:

• Text: a parcel arrived on time with broken glass inside.
• Image: a locally generated 224×224 red PNG; questions ask about its color and red coverage.
• Audio: a 3.5-second, 16 kHz mono PCM clip, synthesized locally with the macOS Samantha voice reading the parcel sentence.

Both models select the damage team for text and red for the image. Omni misanswers two speech questions: it selects book as the damaged item and gives arrival a probability below 0.5. The original implementation produces the same answers. The audio result establishes native/reference agreement on this clip, not speech-task accuracy.

Reproduction

Place the pinned original checkpoint assets in causal/ and omni/ under a local directory. The native test uses the committed input media and independent reference answers:

MERERUN_TEST_MLX_DEVICE=gpu \
MERERUN_TEST_D1_RELEASED_CHECKPOINTS=/path/to/original-checkpoints \
  swift test --filter D1ReleasedCheckpointTests

The regression checks all probabilities, confidence, expected ordinal scores, selected choices, answer types, and input/output-token counts. FP32 absolute tolerance is 0.0001; BF16 absolute tolerance is 0.02. The test skips when the explicit checkpoint directory is absent. Normal user installation remains model pull followed by text decide; the test fixture directory structure is only for qualification.

To regenerate expected answers, retain the original pinned Python source files alongside each checkpoint and run scripts/fixtures/export-d1-release-reference.py for each family. The fixture README records exact dependencies and commands.

Local gates

All 13 focused D1 tests passed with GPU execution, both real checkpoints, original tokenizers, and complete text/image/audio smoke inputs: zero skips or failures. ./scripts/check.sh passed lint, architecture/package-policy checks, build, 5,349 XCTest cases (443 conditional skips, zero failures), 191 Swift Testing cases, CLI help, and hygiene scans. The docs site build also passed.

Scope

This qualifies original-checkpoint loading and agreement for these five requests. It does not establish benchmark accuracy, latency, maximum-context behavior, all-input BF16 equivalence, a tiled/multiple-image checkpoint suite, or a broad audio evaluation. Small independent tensor, resize, and tokenizer fixtures cover additional component behavior. The upstream tree-packing optimization remains unimplemented.