colibri

• #1577

  [Bug]: qwen36 CUDA_DLL on Windows: tier selects 0 devices unless COLI_GPUS is set (available_device_count is queried before the DLL is loaded)

  更新于 2026年9月16日
• #1370

  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

  更新于 2026年9月16日
• #1576

  glm53: KV prefix reuse lost after bare tool-call turns — template writes "\n<tool_call>", the model does not

  更新于 2026年9月16日
• #1498

  [Bug]: GPU doesn't engage

  更新于 2026年9月16日
• #1572

  [Feature]: Adopt Git LFS for test fixtures, large dataset artifacts, and media assets

  更新于 2026年9月16日

核心特点

• •One hierarchy, not limited by tier capacity. VRAM, RAM, and NVMe are placement
• •A JIT for weights. Measured routing heat drives a per-layer LRU, a learned
• •I/O is part of the engine. Batched expert unions, overlapped reads and
• •Heterogeneous execution. CPU, CUDA, Metal, NUMA memory, and partial or full
• •Compressed state without a different model. Token-exact forward validation,
• •Speculation that must earn its keep. Native MTP and grammar-forced drafts
• •the dense part (attention, shared experts, embeddings — ~17B params) stays
• •the 19,456 routed experts (75 MoE layers × 256 + the MTP head, ~19 MB each