llama.cpp adds a native /v1/systemone endpoint for open decision models
llama.cpp merged PR #29818 on 2 October 2026 (06:56 BRT), adding a native /v1/systemone endpoint to llama-server. It covers five open decision models: Laya, Julia-1, Lev, OpenJev 27B and Kev. ggml-org publishes official GGUFs for all five.
Correction (3 Oct). The first version of this note linked “OpenJev” to openjev, the MIT NLI stack by AlexWortega. The model llama.cpp supports is a different one: openjev/openjev, a 27B fine-tune of Qwen3.8-27B under CC BY-NC 4.0 (non-commercial). It now has its own page.
Georgi Gerganov’s announcement (11:33 BRT) drew more than 1,100 likes and 39,000 views in a few hours. The same day, Xuan-Son Nguyen and Victor Mustar published the ggml-org blog post “New in llama.cpp: Decision Models” (@ngxson’s post, 11:25 BRT). It shows llama serve -hf ggml-org/<model>-GGUF, router mode for several models behind one server, and tips such as trying different quantizations. Its latency figures are UnverifiedClaim: median per question on one RTX PRO 6000, measured by the authors.
| Model | Base | Median latency (blog, UnverifiedClaim) | GGUF |
|---|---|---|---|
| Julia-1 | 144M mmBERT-small encoder | 3 ms | ggml-org/Julia-1-GGUF |
| Laya | 421M ModernBERT-large encoder | 5 ms | ggml-org/Laya-GGUF |
| Kev-4B | Qwen3.5-4B-Base | 12 ms | ggml-org/Kev-4B-GGUF |
| Lev | Qwen3.5-4B | 36 ms | ggml-org/lev-GGUF |
| OpenJev 27B | Qwen3.8-27B, text + images, CC BY-NC 4.0 | 43 ms | ggml-org/OpenJev-GGUF |
Clef. Support for Clef came in a separate PR, #29831, merged on 2 October at 21:50 BRT (text only, one sequence per batch). The blog had listed it as “next”. PR #29832, a model-agnostic implementation, is still open.
What changes on the site. Serving through llama.cpp’s /v1/systemone counts as an integration (see the quadrant method). Maturity goes up for three models: Julia-1 4.3 → 4.8, Lev 5.2 → 5.6, Kev 6.1 → 6.4. Clef stays at 7.2, since llama.cpp alone does not move its integration score. Laya’s maturity was recomputed with the weighted formula and is now 6.2 (it was 7.0, set before that formula). OpenJev 27B enters the catalog at 5.8 / 7.0 / 12.