Kolibri-1 EXL3 K6 (routed experts only)
Derived from Aleph-Alpha/Kolibri-1-BF16,
revision ee7d86e7d6ffdaaeb7fb6bd486f52a1e31cb4ce1.
Original model by Aleph Alpha, distributed under Apache-2.0; see LICENSE.
Only the 57,600 routed expert projections are quantized to EXL3 K6/mul1. All 753 other tensors retain their original BF16 dtype and byte content, including attention, shared experts, router, embeddings, norms and output head. The config's head_dtype=float32 describes computation, not stored weight dtype.
Calibration used all 250 original rows of 2,048 tokens, with zero excluded rows. Conversion and export verification passed. This is not a claim of FP8-equivalent quality, task benchmark parity, or verified Radiance inference support.
Each expert projection uses trellis (I16), suh/svh (F16), and a scalar mul1 (I32). K6 uses a 96-word last trellis dimension. mul1's unsigned bit pattern is 0x83DCD12D. The decoder also requires the EXL3 tile permutation and normalized 128-element Hadamard transforms; this is not ordinary affine INT6.
A compatible EXL3 loader and Kolibri architecture implementation are required. Generic Transformers loading is not asserted to work. Use config.json, quantization_config.json and model.safetensors.index.json for the full inventory.
Radiance container
Kolibri-1-EXL3-K6.rad packages this existing EXL3 K6/mul1 export into a Radiance .rad container without requantization or source weight edits.
Converted with Radiance 1.7.0. Loading requires the matching kolibri1/exl3 architecture and EXL3 kernels; this is not a generic upstream drop-in container.
CPU-only conversion and the full-artifact exact source/container audit passed before upload. The audit checks packed expert planes and retained nonexpert tensor bytes; this is not an inference or model-quality evaluation.
This upload does not establish full-model inference compatibility. Full-engine inference, benchmark parity, and general loader compatibility are not claimed. The original Safetensors export and its configuration files remain available unchanged.
Container size: 62241273856 bytes.
Container SHA-256: d652b00c6cd34beb1909401d94d224b914e32ecba56a194b1e89023666a8a769.
- Downloads last month
- 26
Model tree for Code0000/Kolibri-1-EXL3-K6
Base model
Aleph-Alpha/Kolibri-1-BF16