Corrado Nicola Orazi Barattieri

Pull requests

• oMLX pull request #2161

  Merged 2026-09-16

  Calibrate models larger than RAM: the quantiser streams the model one layer at a time. About 3,000 lines.

• oMLX pull request #2159

  Merged 2026-07-10

  Fix the decoding of MXFP8 block scales in dequantisation.

• oMLX pull request #2140

  Merged 2026-07-09

  Ship the calibration data inside the package.

• oMLX pull request #3504

  Open since 2026-09-07

  Grammar-constrained output that keeps speculative decoding on.

Issues

• oMLX issue #2238

  Closed 2026-09-24

  Mixture-of-experts decoding ran at 34% of the memory roofline, against 94% for a dense model, on an M3 Ultra.

• oMLX issue #3507

  Closed 2026-09-12

  The embeddings endpoint answered 200 with vectors full of nulls when the model returned NaN.

• oMLX issue #3505

  Open since 2026-09-07

  The memory guard sets custom tiers to a soft limit of 0.85 and gives no setting for prefill headroom.

• oMLX issue #3506

  Open since 2026-09-07

  The native 8-bit affine kernel for Qwen3.5 cannot engage during prefill, because of its threshold against the chunk size.

• Omi pull request #7092

  Merged 2026-06-10

  Deduplicate when a staged task is promoted.

• Omi pull request #7093

  Merged 2026-06-09

  Make creating an action item idempotent with a content hash.

• Omi pull request #7091

  Merged 2026-06-05

  Sync: adopt orphan action items by description.

• mlx-embeddings #80

  Open since 2026-09-07

  A bug report. Padded ModernBERT batches return NaN embeddings when the longest sequence is a multiple of 32 tokens, at 128 tokens or more.

• Nimbalyst #1161

  Open since 2026-08-05

  Adds Kimi Code as an agent provider.