Issues

• SAM 3 lite doesn't mask padding in the text prompt, so its scores differ from Sam3Model

  apple/coreai-models#381 · Filter by author reda-i opened on Oct 25, 2026
  1

  

• [Swift runtime] Constrained generation traps with "Storage size insufficient for tensor layout" on the first multi-token jump-forward step when the bundle has a prefill graph (logits buffer holds one row)

  apple/coreai-models#380 · Filter by author iannellomarco opened on Oct 24, 2026

• [Swift runtime] Schema-constrained generation skips the prefill graph on bundles without a sliding ring, so the logits buffer grows with the prompt (+0.5 to 0.8 GB at 3-4K tokens on a 100K vocabulary)

  apple/coreai-models#379 · Filter by author iannellomarco opened on Oct 24, 2026

• llm-runner --raw-tokens with --save-logits scores tokens[1...] against tokens[0] instead of generating

  bug
  apple/coreai-models#373 · Filter by author Lewis300 opened on Oct 23, 2026

• Apple 🍎 accounts unlock

  apple/coreai-models#372 · Filter by author mdaliahmedpiku-cpu opened on Oct 23, 2026

• ModelScope as an alternative model download backend

  apple/coreai-models#371 · Filter by author yrk111222 opened on Oct 23, 2026
  2

• [Python runtime] Repeated calls to one function abort the process after ~3000 calls with "Failed to allocate storage for NDArray" (GPU preferred; CPU preferred is fine)

  apple/coreai-models#366 · Filter by author reda-i opened on Oct 22, 2026
  1

• [Tracking] Embedding Model Runtime

  apple/coreai-models#361 · Filter by author stikves opened on Oct 21, 2026

• [Tracking] Batched inference integration

  enhancement
  apple/coreai-models#360 · Filter by author stikves opened on Oct 21, 2026
  1

  

• Model request: FastVLM-0.5B, FastVLM-1.5B, FastVLM-7B

  apple/coreai-models#353 · Filter by author tmorales2000 opened on Oct 16, 2026

• InfiniteTalk

  apple/coreai-models#343 · Filter by author stikves opened on Oct 11, 2026

• Streaming detokenization is O(N²): it re-decodes the whole token sequence every step

  apple/coreai-models#341 · Filter by author stikves opened on Oct 11, 2026
  1

• Aleph-Alpha/Kolibri-1

  apple/coreai-models#330 · Filter by author stikves opened on Oct 6, 2026

• Model request: Speech enhancement — RNNoise, GTCRN and UL-UNAS

  apple/coreai-models#325 · Filter by author maxffarrell opened on Oct 5, 2026

• Model request: Julia-1 decision model for macOS

  apple/coreai-models#324 · Filter by author maxffarrell opened on Oct 5, 2026

• [Swift runtime] Streamed text re-emits text when a token extends the previous grapheme (ZWJ emoji, combining marks, Indic signs)

  apple/coreai-models#315 · Filter by author iannellomarco opened on Oct 1, 2026

• Core AI GPU gather from a constant with 2^31 or more elements silently returns wrong rows; exporter should warn or split (affects gemma-3n-E4B per-layer embeddings)

  apple/coreai-models#313 · Filter by author iannellomarco opened on Oct 1, 2026

• [Swift runtime] A failed GPU command buffer is swallowed: the call returns normally and the pipelined engine streams <pad> tokens

  apple/coreai-models#312 · Filter by author iannellomarco opened on Oct 1, 2026

• Sliding-window SDPA (is_causal + window_size) on strided K/V views returns NaN for a short chunk over more than 3,584 keys in deep dynamic graphs

  apple/coreai-models#311 · Filter by author iannellomarco opened on Oct 1, 2026

• [Python runtime] The 7th distinct input shape of a dynamic function fails (MTL4CommandQueueErrorDomain error 1), returns zeros, and the function stays broken

  apple/coreai-models#310 · Filter by author iannellomarco opened on Oct 1, 2026

• [Swift runtime] Guided generation with Qwen3-4B repeats until maximumResponseTokens; timeout cancellation not honoured (experimental export)

  apple/coreai-models#300 · Filter by author emusickid opened on Sep 29, 2026

• Tag a release that builds with Xcode 27

  bug
  apple/coreai-models#293 · Filter by author willswire opened on Sep 26, 2026

• Transformers version

  apple/coreai-models#292 · Filter by author kevchengcodes opened on Sep 26, 2026

• Explore Video Diffusion VAE tile size

  bug
  apple/coreai-models#285 · Filter by author kevchengcodes opened on Sep 22, 2026

• The RoPE cos/sin gather composite allows only one lookup per graph, fed straight from the position input

  apple/coreai-models#280 · Filter by author baozzz1 opened on Sep 22, 2026