Cohere's 2.4B multilingual VLM converted to Apple's Core AI (the Core ML successor announced at WWDC26), running on iOS 27 and macOS 27. On an iPhone 17 Pro it answers about an image at 18.2 tok/s and 24/24 tokens identical to the fp32 reference — not "close", identical.

Two bundles, run in sequence: a 400M vision tower (custom-trained from SigLIP2-SO400M, and structurally a Qwen3-VL visual encoder — deepstack mergers included) emitting image_embeds [256, 2048] + deepstack_embeds [768, 2048], and a 2B Cohere decoder with the image tokens spliced in as extension ids. Eleven languages, including Japanese. Apache-2.0.

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 Mac tok/s row is the text core because llm-runner cannot bind an image buffer. The iPhone rows are PipelinedBench, including a mandatory 1024-token generation (the iOS compiler miscompiles KV specializations at seq ≥ 2048 and a 256-token run cannot see it) — clean.

int8 costs this model nothing measurable. int4 is not published: 0 of 9 cases, with a lost sentence boundary and instruction boilerplate leaking in ("…a blanket.Answer: Cats.I apologize, but I cannot provide a detailed description of the image") plus flat repetition. int4 tolerance is a property of the individual model — a sibling port at 3B took int4 for free and a 450M one cratered — so read the generations rather than the parameter count.

The coreai-pipelined-static-inputs patch (which binds image_embeds, deepstack_embeds and the two rope-shift scalars) is in the zoo under apps/.

For the image path the host resizes to a 512×512 canvas with an antialiased BICUBIC filter, normalizes (x/255 − 0.5)/0.5, and patchifies into 16×16 patches in Qwen-VL order — block-major over 2×2 merge groups, and [C][T][py][px] inside each patch with the still frame duplicated (patch_dim 1536). Then it runs the tower, binds both outputs, and rewrites the prompt's <image> ids (255031) to V + slot. Reference implementation: _smoke/northmv_preprocess.py.

Note the export bakes a square grid while the upstream processor is native-resolution (it keeps a 640×480 image's own 30×40 patch grid), so non-square images are stretched.

The oracle needs transformers git main: the 5.15.0 release does not know cohere_compass and raises on AutoConfig.

The decoder is where the work is, and each of these runs fine when done wrong: a parallel block (one LayerNorm, attention and MLP summed into the residual), Cohere LayerNorm (the mean is subtracted, no bias), SSSF × 7 layer types where the 7 full-attention layers have no positional encoding at all while the 21 sliding ones carry interleaved M-RoPE in a 4096 window, logit_scale 0.25, and a 262 144-entry embedding tied to the head.

The tower needed no work: it loads into the zoo's existing Qwen3-VL encoder with zero missing keys and matches every seam at cos 1.000000.

Everything is in conversion/export_northmv_pipelined.py and knowledge/north-micro-vision-port.md.

Apache-2.0, carried from CohereLabs/North-Micro-Vision-Instruct (revision 373bda96ac70bf89f99f7048f420cf00dc07c149). Not affiliated with Apple or Cohere.

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.