english-monlam-local-private

Experimental public encoded weights for the New Tibet private inference runtime. This is a CKKS execution package, not a Transformers checkpoint or hosted API. The client owns the secret key, attention, KV cache and nonlinear operations. The community host receives ciphertexts and public evaluation keys only. Model files contain public coefficients; user data and runtime keys are excluded.

Derived from the pinned Monlam Gemma 2 q4 base and English LoRA. Native uint32 Metal arithmetic uses N=8192 and Q/P primes below 2^31. A measured complete-model cached step took 47.27 seconds on an M3 Mac, versus 134.47 seconds for the previous FHE baseline; native English ran at about 13 tokens/second. This is experimental, not an audited malicious-host protocol or a long-context quality evaluation.

Host model data is about 50 GiB on disk and streamed through bounded RAM buffers. Public client embedding/tokenizer assets are separate. See the New Tibet source repository for the matching runtime, benchmark report and exact revisions.

Lineage and modifications

This package merges the pinned Monlam Melong q4 base with the published English LoRA before preparing its FHE projection weights. Both direct parents are listed in base_model so Hugging Face can show this release in their model trees. The FHE representation itself is a format conversion, not additional training.

  • Monlam base: TenzinGayche/Monlam_Melong_preview at 663b1e876362d51324bbd4ac8e7c77fecf4f90f4.
  • English adapter: AlexJupiter/monlam-melong-mlx-english-adapter-preview at b6bb4422ccf841e4a01a8d7cdb42fef325a02430.
  • Runtime source: https://github.com/AlexJupiter/new-tibet-app-testnet
  • Model card and manifest identify modifications to all exported derived assets.

License and notices

This is a Gemma-derived model. Use and redistribution are governed by the included Gemma terms, their incorporated prohibited-use policy, and Notice. By using or redistributing this model, you agree to these terms and must pass the applicable restrictions and notices to recipients. The publisher confirmed redistribution permission and that no additional Monlam notice is required. This package does not relicense upstream weights under the FHE runtime's software license.

The documentation files are included in the checksum-pinned manifest. Runtime secret keys, evaluation keys, chat data and benchmark fixtures are excluded.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for AlexJupiter/english-monlam-local-private