Spaces:

---

multimodalart

/

jev-decision-index

Like 412

Running

Hi! Could you add Jebadiah 27B to the index? It's our best one so far.

• Weights: https://huggingface.co/frontier-infra/jebadiah-27b (Apache-2.0, merged bf16 safetensors, about 56 GB)
• Code, trainer and server: https://github.com/getainode/jebadiah
• Base: Qwen/Qwen3.8-27B (the chat checkpoint, thinking off) with a merged rank-16 LoRA on every linear projection, no extra heads. Trained on public data only.
• Types: choice, noul and score. The probabilities come from the option-label token logits in one prefill, with zero output tokens.
• Languages: English.

It serves the TypeSafe /v1/systemone format:

git clone https://github.com/getainode/jebadiah.git && cd jebadiah/server
uv sync --extra cuda
uv run jebadiah-serve --model frontier-infra/jebadiah-27b --host 0.0.0.0

That server caps choice at 20 options, the same cap AINode serves. The index has bigger menus than that, so for a board run the in-process engine in apolinario/decision-index PR 8 is the better path. I'm updating that PR from the 9B to the 27B now. The same engine is running the full 0.2 suite on our 9B v2 on a DGX Spark right now, ToolRet included, so the big menus are covered.

One RTX PRO 6000 has room for it in bf16. I don't have a run-time estimate for the 27B on your card yet. It will be slower than a 9B.

Our own numbers on the public suites are on the card; they are not board rows. Happy to help if anything breaks on your side. Thanks for running the index!