Issues

• Hermano lenguajes

  JustVugg/colibri#2580 · Filter by author illoquediC opened on Nov 3, 2026

• [Performance]: Qwen3.8-Flash-Next UD-Q2_K_XL on 2x RTX 3090 via Vulkan (v2.1.0): 37-49 tok/s on one card; Q38_MTP=1 slower; COLI_VK_DEV2=auto drops decode to 2-3 tok/s (8592 expert uploads fail)

  JustVugg/colibri#2556 · Filter by author bokiko opened on Nov 2, 2026
  1

• api: _qwen_tool_calls may drop prefix reuse when content ends with a single newline

  JustVugg/colibri#2544 · Filter by author dcoldeira opened on Nov 2, 2026
  1

• [Feature]: can glm5.3? windows gpu nvidia release dll?

  JustVugg/colibri#2541 · Filter by author blstdmi opened on Nov 1, 2026
  8

• [Bug]: GLM-5.3 (glm_moe_dsa) gets the GLM-5.2 prompt without reasoning_effort: no effort line, closed <think></think>

  JustVugg/colibri#2540 · Filter by author mbalser opened on Nov 1, 2026
  1

• [Feature]:

  JustVugg/colibri#2498 · Filter by author maxim333555777w-ship-it opened on Oct 30, 2026

• [Bug]: libomp not found when using MacPorts

  JustVugg/colibri#2429 · Filter by author GabZach opened on Oct 27, 2026
  1

• [Performance]: macOS 27 halves Metal throughput on M1 Max — METAL-ATTN gpu-sched goes 2.06s → ~31s on the same binary, same model, same flags

  JustVugg/colibri#2426 · Filter by author Toao27 opened on Oct 27, 2026
  3

• [Bug]: [glm53][Windows/Vulkan] Every activation NaN from the first MoE layer, any prompt — a first-frame sync race the KDA recurrence then keeps alive

  JustVugg/colibri#2355 · Filter by author DesairDHawk opened on Oct 24, 2026
  3

• Datapoint + setup bug: Qwen3.6-35B-A3B on a pre-AVX2 dual Xeon (Ivy Bridge-EP), CPU-only, HDD RAID — 1.30 tok/s rotating; wizard saves cap 8 despite planning 256

  JustVugg/colibri#2337 · Filter by author CtzarGM opened on Oct 23, 2026
  1

• [Feature]: Embedding support for Docmost integration

  JustVugg/colibri#2317 · Filter by author stefhimself opened on Oct 23, 2026
  1

• [Bug]:Landing page forces a 1280px viewport on mobile devices

  JustVugg/colibri#2294 · Filter by author sparshsingh12312 opened on Oct 22, 2026
  1

• [Bug]: Vulkan engine crashes (vkAllocateMemory -2, exit 0xC0000409) on prompts >= 19K tokens — no chunking before dispatch

  JustVugg/colibri#2288 · Filter by author 364328984 opened on Oct 22, 2026
  2

• Qwen3-Coder-30B-A3B-colibri-int4 checkpoint incomplete: 43/48 layers, missing global tensors and metadata

  JustVugg/colibri#2287 · Filter by author Risingimpacta opened on Oct 22, 2026
  6

• [Performance]: DeepSeek V4 on Windows: libgomp never spins at barriers, so the launcher's spin settings do nothing; a clang + libomp build decodes 1.5x faster with identical output

  JustVugg/colibri#2286 · Filter by author sadanbermudez opened on Oct 22, 2026

• [Performance]: Qwen3-Coder-30B-A3B on Android (Termux), Samsung Galaxy Tab SM-X920 / Dimensity 9300+, 12 GB — 0.33 tok/s CPU

  JustVugg/colibri#2283 · Filter by author javieratcservicios opened on Oct 22, 2026

• I have an issue

  JustVugg/colibri#2275 · Filter by author Hossein-Rzv opened on Oct 21, 2026
  1

• [Performance]: Qwen3.6-35B on GTX 1650 Max-Q (4 GB) + 32 GB laptop: #2098 tier starts, but slower than CPU-only; GPU output not reproducible at temp 0

  JustVugg/colibri#2211 · Filter by author javieratcservicios opened on Oct 18, 2026
  7

• Windows + qwen4exp GGUF: two import bugs and decode ~5x slower than llama.cpp on the same file

  JustVugg/colibri#2189 · Filter by author guo-kewei opened on Oct 18, 2026
  3

• [Performance]: Qwen3.8-27B (dense) on RTX 5060 Ti via Vulkan: 12.0 tok/s at int4, 1 vs 2 GPUs at int8

  JustVugg/colibri#2127 · Filter by author ibboucco opened on Oct 16, 2026
  4

• [Feature]: Possibility of adding the new openbmb/MiniCPM-V-4.7-35B-A3B , LiquidAI/LFM2.5-24B-A2B along with kwanhee/Kimi-K2.5-REAP50-INT4-W4A16, drawais/Nemotron-3-Super-120B-A12B-REAP50-W4A16 if possible

  JustVugg/colibri#2107 · Filter by author gggff123 opened on Oct 15, 2026

• [Performance]: Running on a Sandy Bridge CPU based laptop

  JustVugg/colibri#2106 · Filter by author florin65 opened on Oct 15, 2026
  4

• [Performance]: Benchmark: Kimi K3 (1.56 TB) on Ryzen 9 9950X3D + RTX 5080 (Windows 11, CUDA) + 128 GB DDR5-6000 + D:\Samsung 990 EVO Plus 2 TB (PCIe 4.0) <==> C:\Samsung 9100 PRO 2 TB (PCIe 5.0),

  JustVugg/colibri#2091 · Filter by author DenisZDV opened on Oct 14, 2026
  5

• [Feature]: GPU tier disabled on 32 GB RAM + 4 GB VRAM; trunk placement could still help

  JustVugg/colibri#2090 · Filter by author Yva1n opened on Oct 14, 2026
  5

• Secure browser authentication for the local dashboard on API-key-protected servers

  JustVugg/colibri#2063 · Filter by author trigger2k20 opened on Oct 13, 2026
  1