TurinTech publishes Artemis GGUF builds of eight open decision models that run on stock llama.cpp

Source · Model card: pplx-decider v1.1 27B

From 12 October 2026 TurinTech has been publishing Artemis GGUF builds of open decision models on Hugging Face: pplx-decider 27B, Clef, Clef-Flash, Kev 4B and Kev 9B, then Kev 27B, d3-flash and d3 27B on 14 October. They are conversions, not new models, and TurinTech says it is not affiliated with the original authors. Licences follow the originals (Apache-2.0 on the ones we read).

The files run on stock llama.cpp (builds from 9 October or newer) behind llama-server’s /v1/systemone. The card warns to keep -ub 4096 -b 4096: at the default 512, /v1/systemone rejects inputs longer than 512 tokens. For pplx-decider, TurinTech also ships a vision projector converted from the original weights, because Perplexity does not publish one for llama.cpp.

TurinTech’s numbers for pplx-decider 27B (UnverifiedClaim): the default file is 15.28 GB against 53.8 GB for BF16 and needs 16.6 GiB of GPU memory against 50.8 GiB. It gives the same answer as the original on 96.8% of 2,000 test decisions from typed-decisions, with 70.8% accuracy against 70.5%, and changed no answer the original was more than 80% sure of on an A100. A plain Q4_K_M agrees 95.7% of the time.

Catalog pages: pplx-decider · Clef · Kev · Decision 3.0.