# Checkpoint mirror

This package mirrors the complete official RecursiveSI Flint INT4 target checkpoint at revision `3b93f9053b530488405ae9fca51d4900d6ce21f0`. Recursive is the model author. Fifteen upstream files, including the weights, tokenizer, chat template, configuration, Apache-2.0 LICENSE and logo, are preserved unchanged. README.md is an EigenLabs-authored model card; the upstream README is preserved byte-for-byte as ORIGINAL_README.md (SHA-256 8041980cec84805c27acd78aa918f3e9d8cadeb0e1668e0841abacb5aea225ac). The README is not a runtime file, so the native runtime aggregate is unchanged.

Upstream: https://huggingface.co/RecursiveSI/flint-int4/tree/3b93f9053b530488405ae9fca51d4900d6ce21f0

The artifact uses mixed-bit affine quantization (INT4 projections/experts, INT8 embedding); it is not NVFP4. It is a text-only, target-only package. No compatible published MTP companion is included. The model chooses when to reason; this checkpoint does not provide trained forced thinking on/off, effort-level or thinking-budget controls.

Native d-inference support credits Prince Canuma / mlx-vlm (MIT) and Recursive / rec (Apache-2.0) in the SDK ThirdPartyNotices/Flint.md, with pinned source revisions and complete notices. Their code is not copied into this model-weight mirror.

This mirror is not a Darkbloom catalog release or a performance qualification. No throughput guarantee is made.