Issues

• Reuse reconstructed DFlash rings for repeated resident-prefix requests

  incoai/splash#671 · Filter by author linson007 opened on Nov 4, 2026

• Reproduce complete-edit prompt-lookup gains with portable native benchmarks

  incoai/splash#670 · Filter by author linson007 opened on Nov 4, 2026

• Reconsider guarded prompt lookup with complete-edit DFlash2 measurements

  incoai/splash#669 · Filter by author linson007 opened on Nov 4, 2026

• Admit smaller lazy branch junctions when planned draft work is unchanged

  incoai/splash#667 · Filter by author linson007 opened on Nov 4, 2026

• Guide first-time users from model download to their first chat

  incoai/splash#665 · Filter by author linson007 opened on Nov 3, 2026

• App: Detect and use existing server

  incoai/splash#657 · Filter by author Lancelotbronner opened on Nov 3, 2026

• Would you take an opt-in --log-token-ids diagnostic flag?

  incoai/splash#655 · Filter by author airiclenz opened on Nov 2, 2026

• M5 Pro M16 decode follow-up: +5.92% paired result and storage-safe integration

  incoai/splash#652 · Filter by author linson007 opened on Nov 2, 2026

• Support for 0xKitkat's ultra lite Qwen 3.8 Flash Next

  incoai/splash#647 · Filter by author Lancelotbronner opened on Oct 31, 2026
  2

• LoRa support?

  incoai/splash#646 · Filter by author tomchk opened on Oct 31, 2026

• KV page pool never evicts idle sessions — decode collapses from ~97 to ~12 tok/s on long multi-session use

  incoai/splash#644 · Filter by author sunnybluesea opened on Oct 31, 2026
  6

• Interest in a local AI assistant built on Splash? Early Puddle experiment

  incoai/splash#639 · Filter by author linson007 opened on Oct 30, 2026

• Support for NVIDIA Nemotron 3.5 Bolt (target, LiLiCorr draft, vision)

  incoai/splash#633 · Filter by author ansoncat opened on Oct 30, 2026
  3

• Do we support Qwen4 from Day 0 ?

  incoai/splash#629 · Filter by author linson007 opened on Oct 29, 2026
  2

• Feature request: video input

  incoai/splash#625 · Filter by author elias-fox opened on Oct 29, 2026
  4

• [1.5.0 macOS] Chat input loses CJK IME after switching models / starting new chat (pinyin typed as literal text)

  incoai/splash#623 · Filter by author deyi1031 opened on Oct 29, 2026
  1

• Feature request: run a single model across multiple Macs over Thunderbolt

  incoai/splash#607 · Filter by author linson007 opened on Oct 27, 2026
  1

• Expert streaming keeps the SSD saturated with --max-memory: prefill expert hit rate 36%, ~7 GB/s of reads

  incoai/splash#598 · Filter by author prakash0207 opened on Oct 26, 2026
  1

• Support Qwen3.8-Flash-Next MLX checkpoints

  incoai/splash#597 · Filter by author prakash0207 opened on Oct 26, 2026

• feat: Support Nanbeige4.3-33B

  incoai/splash#595 · Filter by author reneleonhardt opened on Oct 26, 2026
  4

• Missing safeguards field passthrough breaks Claude Code auto mode server-side classifier

  incoai/splash#589 · Filter by author sambengtson opened on Oct 25, 2026
  1

• Resumed prompts re-read up to 36 cached tokens because the replay state sits on a 32-token page

  incoai/splash#588 · Filter by author ghisguth opened on Oct 25, 2026
  5

• Use smaller SIMD-group rows for short Qwen3.8-27B prefill on the 20-core M5 Pro

  incoai/splash#582 · Filter by author linson007 opened on Oct 24, 2026
  1

• Qualify one-lane AF4G64 M16 decode on the 20-core M5 Pro

  incoai/splash#581 · Filter by author linson007 opened on Oct 24, 2026
  2

• Qwen3.8-27B on a 48 GiB M5 Pro: reproducible inference measurements and tuning results

  incoai/splash#580 · Filter by author linson007 opened on Oct 24, 2026
  2