Flint UD-Q4_K_XL, NInfer v3 artifact This repository re-encodes the weights listed below into NInfer's container format. Each source is available under the Apache License 2.0: https://www.apache.org/licenses/LICENSE-2.0 Modifications: none to the weights' values. Every ggml block of the GGUF is copied without requantization; Q, K and V of a block are stored as one parent when the GGUF stores them in one format, as are the dense MLP's and the shared expert's gate and up rows, and each routed expert's gate rows are followed by its up rows. The BF16 router rows are widened to FP32, which holds every one of their values exactly; norms and attention sinks are the GGUF's FP32 values. The tokenizer, chat template and generation config are Recursive's files. --- Flint (RecursiveSI/flint, revision 2a247be0): model, tokenizer, chat template, generation config --- Recursive, 2026, Apache License 2.0. --- Flint GGUF, UD-Q4_K_XL (unsloth/flint-GGUF, revision 20468edf) --- Unsloth, 2026, Apache License 2.0. flint-UD-Q4_K_XL.gguf, sha256 63c2ecc461ca545134b670c660e76fa20f47240f8d39b6510bfd21c4dd6e072f.