Issues

• [Performance]: Qwen3.8-Flash-Next FP8 on the Vulkan expert tier, RTX 4070 Laptop 8 GB (Windows 11)

  JustVugg/colibri#1900 · Filter by author namespaceMarcello opened on Oct 4, 2026

• [Feature]: New Model Kolibri-1

  JustVugg/colibri#1869 · Filter by author MichaelFomenko opened on Oct 3, 2026

• Issue when trying to use Colibri with OlMoE

  JustVugg/colibri#1866 · Filter by author MechaLogicEngineer opened on Oct 3, 2026
  1

• test_inkling_shared_batch: "bf16 mode=default is not scalar bit-exact" on an AVX2-only host (make check fails locally, CI green)

  JustVugg/colibri#1864 · Filter by author crichalchemist opened on Oct 3, 2026
  2

• glm53: "dense matrices on the device" with f32 dense matrices, but none are ever dispatched (vulkan_engines.sh glm fails on a discrete GPU)

  JustVugg/colibri#1863 · Filter by author crichalchemist opened on Oct 3, 2026
  2

• tests/vulkan_engines.sh: qwen, inkling-olmoe and mimo-qwenimage fail on a discrete GPU now that the dense chain is on by default there

  JustVugg/colibri#1862 · Filter by author crichalchemist opened on Oct 3, 2026
  2

• [Feature]: Deb or pkg

  JustVugg/colibri#1852 · Filter by author dreadbughunter opened on Oct 3, 2026
  12

• [Feature]: Asynchronous PIPE queue

  JustVugg/colibri#1810 · Filter by author tonnthuir opened on Sep 30, 2026
  1

• [Feature]: Zero Weight Caching Mode

  JustVugg/colibri#1764 · Filter by author MichaelFomenko opened on Sep 26, 2026
  3

• [Feature]: improve glm53 cache reuse after interruption

  JustVugg/colibri#1748 · Filter by author enitimeago opened on Sep 25, 2026
  2

• Ternary weight formats (~1.75-2.13 bpw): in scope? and how to handle the GGUF/safetensors gap

  JustVugg/colibri#1703 · Filter by author kiljoy001 opened on Sep 23, 2026
  3

• [Feature]: close the OpenAI-API compatibility gap in c/openai_server.py, in value tiers

  JustVugg/colibri#1698 · Filter by author monotophic opened on Sep 23, 2026
  3

• [Feature]: Disaggregated prefill

  JustVugg/colibri#1672 · Filter by author ggeorgovassilis opened on Sep 22, 2026
  3

• [Performance]: macOS planner trusts free + inactive, but the kernel compresses the expert cache — 32 GiB budget is slower than 16 GiB on a 48 GB Mac

  JustVugg/colibri#1614 · Filter by author Freizeitminister opened on Sep 18, 2026
  4

• [Bug]: GPU doesn't engage

  JustVugg/colibri#1498 · Filter by author Namekinek0 opened on Sep 14, 2026
  6

• [Feature]: Provide the necessary Windows coli_cuda.dll files or update instructions to build them

  cudafeature
  JustVugg/colibri#1405 · Filter by author SoftologyPro opened on Sep 9, 2026
  20

• Perplexity vs Ollama's Q4_K_M on the same tokens: our int4 experts lose 2.5–3.4 %, the int8 trunk buys nothing — two questions about the format

  JustVugg/colibri#1370 · Filter by author kreuzzelg opened on Sep 6, 2026
  3

• DeepSeek V4 has no Metal path — measured 5.05x TTFT / 2.81x tok/s gap on Apple Silicon

  JustVugg/colibri#1309 · Filter by author BrianHeeseIs opened on Aug 31, 2026
  3

• [Discussion] Upstreaming the #161 CUDA stack: attention is now 80% of resident-host decode, and the 2.4→11.1 line deserves a continuation

  JustVugg/colibri#1140 · Filter by author ZacharyZcR opened on Aug 20, 2026
  3

• [EXPERIMENT] Prompt-semantic Atlas cluster pinning for cold-start prefetch — pre-registered A/B (README hypothesis row 1)

  JustVugg/colibri#1078 · Filter by author siongchin87 opened on Aug 18, 2026
  7

• [perf] olmoe: O(cap) eviction scan under g_pilot_mx makes tok/s non-monotonic in cache size

  good first issueperformance
  JustVugg/colibri#1050 · Filter by author sami7969 opened on Aug 16, 2026
  7

• [Feature]: Add support for dots-studio/dots3-note-prev

  featuremodel-support
  JustVugg/colibri#1004 · Filter by author Dynoooooo opened on Aug 14, 2026
  1

• [Feature]: Add support for Mistral Large 3

  featuremodel-support
  JustVugg/colibri#972 · Filter by author CorentinWicht opened on Aug 12, 2026
  3

• # [Feature] Metal: tiled tensor-ops prefill GEMM — int8 tensor ops now (26.5: W4A8/W8A8), fp8-native next (macOS 27) — and routed experts are still CPU at prefill

  featurehardware-owner-neededmetalperformance
  JustVugg/colibri#918 · Filter by author RDouglasSharp opened on Aug 10, 2026
  1

• OLMoE has no hot expert zone (counter-datapoint for pinned hot-store policies)

  benchmarkdiscussionperformance
  JustVugg/colibri#864 · Filter by author outtodata opened on Aug 6, 2026
  5