LFM2.5-VL-3B — Apple Core AI (`.aimodel`)

LiquidAI's LFM2.5-VL-3B converted to Apple's Core AI (the Core ML successor announced at WWDC26), for macOS 27. The detail tier of this family: where the 450M answers "two cats on a pink couch", the 3B answers "the cat on the left is smaller, with a gray and black striped coat, while the cat on the right is larger with a brown and black striped pattern."

Two bundles, run in sequence: a SigLIP2-NaFlex vision tower + projector (patches [1024,768] → image_embeds [256,2048], hidden 1152 × 27 layers) and the LFM2 conv+attention hybrid decoder (hidden 2048, 30 layers = 22 short-conv + 8 GQA attention, vocab 128 000, tied head), with the image tokens spliced in through a static image_embeds input. No recurrent scan, so decode is loop-free on Apple's coreai-pipelined GPU engine with no custom kernels.

Requires macOS 27 (Core AI ships with the OS). Conversion code, gates and knowledge base: coreai-model-zoo.

Bundles

path	size	measured (M4 Max)	numerics
`gpu-pipelined/lfm2_5_vl_3b_vision_fp16`	815 MB	75.7 ms/image	`image_embeds` cos 0.999995 vs fp32 HF
`gpu-pipelined/lfm2_5_vl_3b_decode_int8lin`	3.1 GB	—	suite 7/9 cases token-exact; `logits_last` cos 0.999970
`gpu-pipelined/lfm2_5_vl_3b_decode_int4lin`	2.0 GB	—	suite 7/9 — identical to int8 and to the fp16 baseline
`gpu-pipelined/lfm2_5_vl_3b_decode_int8lin_textcore`	3.1 GB	120.9 prompt / 105.3 decode tok/s	the same weights with no image input

M4 Max, macOS 27.0 (26A5378n), Xcode 27.0 (27A5218g), coreai-torch 0.4.1, llm-benchmark -p 128 -g 256 -n 3, COREAI_CHUNK_THRESHOLD=1. The tok/s row is the text core because llm-runner cannot bind the VLM bundle's image_embeds buffer.

int4 costs this model nothing, which is worth stating plainly because the 450M sibling craters at int4 (0 of 9 cases). Judged against an fp16 baseline rather than fp32 alone — greedy decoding turns any near-tie into a different tail, and the fp16 bundle itself lands 7/9 — int8lin and int4lin both reproduce that 7/9. The divergences are wording: "sleeping peacefully on a bright pink couch" → "sleeping on a pink couch".

iPhone 17 Pro — `ios-h18p/lfm2_5_vl_3b_decode_int4lin` + the fp16 tower

27.5 prefill / 19.3–22.8 decode tok/s, nat 16/16 and image oracle 24/24, clean at a 1024-token generation. int8lin does not load on iOS (its AOT resources.bin is 3.13 GiB); int4lin's is 2.03 GiB and does — which is worth stating because the note this port was written against put the load wall at 2 GiB, and 2.03 GiB was written up as expected-to-fail before a phone was asked. It loaded. Use int8lin on a Mac and int4lin on a phone; on this model int4 costs nothing (7/9 on the suite, the same cases as fp16).

On device the description matches fp32's picture and diverges at the same near-tie the Mac bundles take ("sleeping peacefully" → "sleeping"), then onto an equally accurate branch.

Run it

git clone https://github.com/apple/coreai-models   # + the zoo's engine patches, see below
swift build -c release --product llm-runner

COREAI_CHUNK_THRESHOLD=1 .build/release/llm-runner \
  --model gpu-pipelined/lfm2_5_vl_3b_decode_int8lin_textcore \
  --prompt "The alphabet begins A, B, C," \
  --max-tokens 64 --sampling-strategy greedy \
  --inference-engine-variant coreai-pipelined --warmup off

The engine patches (coreai-pipelined-extra-states for the conv state, coreai-pipelined-static-inputs for image_embeds) are in the zoo under apps/.

For the image path the host resizes to 512×512, normalizes (x/255 − 0.5)/0.5, and patchifies into 16×16 patches with the channel as the fastest axis ([y][x][c]); then it runs the vision bundle, binds the output as image_embeds, and rewrites the prompt's <image> ids (124907) to V + slot. Reference implementation: _smoke/lfm25vl_preprocess.py.

Two host details differ from the 450M and both are silent when wrong. This checkpoint declares resample: 3 (PIL BICUBIC) where the 450M declares 2 (BILINEAR) — read it off processor_config.json. And this tokenizer's post-processor does not prepend <|startoftext|> (the 450M's does), while the chat template starts with it: feed the model a prompt without BOS and it answers " F, F, F, F" — fluent degeneracy, no error.

Converting this family yourself

Build the oracle on transformers ≥ 5: 4.57.6 applies the projector's LayerNorm unconditionally while this config sets projector_use_layernorm: false and ships no such weights, and nn.LayerNorm's default init makes that invisible.

The weight shapes give away the rest: patch_embedding.weight is [1152, 768] — a Linear over pre-flattened patches, not a Conv2d over an image — and position_embedding.weight is [256, 1152], a 16×16 grid bilinearly resized (antialias) to the actual patch grid. The tower's 4304-wide MLP is not divisible by 32, so int8 there is per-block-16.

Everything is in conversion/export_lfm25vl_pipelined.py (--hf-id LiquidAI/LFM2.5-VL-3B — the same script that built the 450M) and knowledge/lfm2.5-vl-port.md.

License

LFM Open License v1.0, carried from LiquidAI/LFM2.5-VL-3B (revision 5a414ead75d45db003906d06fb62bd5b6846cec0). Not affiliated with Apple or LiquidAI.

More models in this format: Core AI Model Zoo — 75 models, each with the recipe that produced it.

Want a different model on-device? Open a request — free, open weights only; the export and its measured numbers get published publicly.