You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The Core AI starter models table (website/src/docs/core-ai/models.md) links community-exported bundles for some models and marks the rest as "Planned Callstack-hosted bundle":
qwen2.5-1.5b-instruct (language)
sana-sprint-0.6b (image generation)
efficient-sam-vitt (image segmentation)
yolos-tiny, yolos-base (object detection)
Export these with the apple/coreai-models 1.0 recipes (iOS presets) and host them, e.g. under a Callstack Hugging Face org, then update the table with direct links. Hosting our own copies also removes the dependency on third-party uploads for the models that are linked today.
The issue requests a new enhancement: exporting and hosting specific Core AI starter model bundles under a Callstack-controlled source and updating the documentation links. It identifies concrete models and deliverables, making it actionable.
Still open: qwen2.5-1.5b-instruct and yolos-base. Both can be exported with the coreai-models 1.0.0 recipes (coreai.llm.export Qwen/Qwen2.5-1.5B-Instruct --platform iOS and models/yolo/export.py --model hustvl/yolos-base), then hosted next to the others.
The hosted FLUX.2 export also needs its image-to-image parts (vae_encoder_half, transformer_512_img2img_half) for #343 (#372). The current coreai-models README says --platform iOS includes them, but the hosted export doesn't have them, so the parts may need --components at 1.0.0.
Activity
grabbou commentedon Oct 26, 2026
Status after the #237 work:
Hosted (exported with the
coreai-models1.0 recipes, tested on an iPhone 17 Pro), and listed on the models page:grabbou/qwen3-0.6b-CoreAI(replaces the third-party link)grabbou/FLUX.2-klein-4B-CoreAIgrabbou/Wan2.1-T2V-1.3B-CoreAI(macOS; iPhone tracked in fix(core-ai): fit video generation in iphone memory #306)grabbou/parakeet-tdt-0.6b-v3-CoreAIgrabbou/clip-vit-base-patch32-CoreAIgrabbou/yolos-tiny-CoreAIgrabbou/EfficientSAM-Ti-CoreAI, video:grabbou/sam3-video-CoreAIgrabbou/edsr-x2-CoreAI,grabbou/pvt-v2-b0-CoreAI,grabbou/depth-anything-3-small-CoreAIStill open from this issue:
qwen2.5-1.5b-instructyolos-basesana-sprint-0.6b: its Gemma-2-2B text encoder is 4.9 GB, which does not fit an iPhone app, so it stays macOS-onlyMoving the hosted exports to a Callstack organization is tracked in #308.
🤖 Generated with Claude Code
[-]Core AI: host starter model bundles[/-][+]chore(core-ai): host starter model bundles[/+]apex-sentinel-staging commentedon Nov 1, 2026
Apex Sentinel · Issue review
Suggested label:
featureThe issue requests a new enhancement: exporting and hosting specific Core AI starter model bundles under a Callstack-controlled source and updating the documentation links. It identifies concrete models and deliverables, making it actionable.
grabbou commentedon Nov 1, 2026
Still open:
qwen2.5-1.5b-instructandyolos-base. Both can be exported with thecoreai-models1.0.0 recipes (coreai.llm.export Qwen/Qwen2.5-1.5B-Instruct --platform iOSandmodels/yolo/export.py --model hustvl/yolos-base), then hosted next to the others.The hosted FLUX.2 export also needs its image-to-image parts (
vae_encoder_half,transformer_512_img2img_half) for #343 (#372). The current coreai-models README says--platform iOSincludes them, but the hosted export doesn't have them, so the parts may need--componentsat 1.0.0.🤖 Generated with Claude Code