We ship Swift/MLX packages that name a weight repo in code. When an upstream repo moves or
disappears, every already-published version of those packages breaks for anyone installing
fresh — there is no redirect and no way to patch software already in someone's hands.
That is not hypothetical: an unrelated model family we had ported was reorganized onto a new
HuggingFace org without redirects, and six of our published packages broke the same day. This
mirror exists so that the same failure cannot take LFM2.5 with it.
It exists for durability, not improvement. Nothing here is changed, tuned, or repackaged,
and LiquidAI's original is the better place to get
these weights if you want the newest revision. All model IP, training, and quantization are
LiquidAI's work.
Integrity
Byte-for-byte identical to upstream. Every file was verified by SHA-256 against the
corresponding LiquidAI/LFM2.5-8B-A1B-MLX-5bit file immediately before upload, and the published copies were
re-verified afterward. The hashes are listed below so the claim is checkable rather than
merely asserted:
Small non-LFS files (configs, tokenizer metadata) carry no upstream hash to compare against and
were verified by exact byte size.
Licence — the commercial threshold matters
These weights are governed by the LFM Open License v1.0, included in full as
LICENSE. Mirroring changes nothing about it.
It is Apache-2.0-shaped, with one condition that is easy to miss and important if you are
building a product rather than experimenting:
Section 5 — Commercial Use limitation. Commercial use by a legal entity with annual
revenue at or above $10,000,000 USD is not licensed under this agreement. Below that
threshold, commercial use is permitted. Qualified non-profits are exempt for non-commercial
and research purposes.
Redistribution is expressly permitted by §4, on four conditions: supply a copy of the licence,
mark any files you modified, retain attribution notices, and carry forward any NOTICE file. We
have modified nothing.
This summary is a pointer, not a substitute for reading the licence, and where the two
differ the licence governs.
Credit
LFM2.5-8B-A1B is designed, trained, quantized, and released by LiquidAI.
This mirror adds nothing but an address we can promise not to move.
