Jeeves-9B

A Jev-like decision model that reasons before it decides when it is unsure, trained on public data only. Give it a state (text or JSON) and yes/no, multiple-choice or rating questions; it answers directly when it is confident, otherwise writes a short reasoning chain first, and returns a calibrated probability for every option.

This is Jeeves v2. v1 (SFT + CISPO) is the v1 tag of this repository: hf download PostHog/jeeves --revision v1.

This repository holds the fused weights (Qwen3.5-9B with the LoRA merged), the pointer head, the two fitted temperatures, and two speculative-decoding drafters. It loads with the Jeeves code, not with transformers: the pointer head and the prompt format are part of the model.

Files

file	contents
model-*.safetensors, config.json	fused Qwen3.5-9B weights, bf16
head.pt	pointer head (query and key projections, 256-dim)
export.json	temperatures (1.522 direct, 5.420 after reasoning), prompt format, training step and config
tokenizer.json, vocab.json, merges.txt, tokenizer_config.json, chat_template.jinja	Qwen3.5 tokenizer
drafter_k4.safetensors	block-4 drafter, the serving default, bf16 (trained on v1; v2 drafters to follow)
drafter_k8.safetensors	block-8 drafter, faster for single-question requests, bf16
drafter_k4.config.json, drafter_k8.config.json	drafter training settings
LICENSE	Apache-2.0, inherited from Qwen3.5-9B

Usage

hf download PostHog/jeeves --local-dir jeeves-weights
git clone https://github.com/PostHog/jeeves && cd jeeves
python -m inference.serve --model ../jeeves-weights --drafter ../jeeves-weights/drafter_k4.safetensors --port 8009

curl -s localhost:8009/v1/systemone -H 'content-type: application/json' -d '{
  "state": "I was charged twice for one order.",
  "questions": {
    "billing": {"type": "noul", "instructions": "Is this about billing?"},
    "tone": {"type": "choice", "instructions": "What is the customer'"'"'s tone?", "criteria": {"calm": null, "frustrated": null, "angry": null}}
  }}'

The request and response follow Jev's /v1/systemone format. An optional options object sets think, max_think (truncate chains), nothink_threshold (skip thinking when the no-think answer is already confident) and return_reasoning. Serving runs on a CUDA GPU or on an Apple Silicon Mac with 48 GB (add --max-rows 4 --max-len 4096 there). It serves bf16 by default; --precision fp8 runs the linear layers in FP8 on GPUs with compute capability 8.9 or higher and on Macs. FP8 weights can be downloaded from PostHog/jeeves-fp8.

Results

Jeeves answers directly when its calibrated confidence is at least 0.9 and otherwise reasons first (nothink_threshold 0.9, greedy, 2,560-token cap); the percentage is the share of questions that reasoned. Kev-9B and Jev numbers are the ones Kev publishes. JevK5 (v0.3) does not reason; we ran it ourselves on the same items. JevBench uses the same 231 public items for every model (the sealed judge tier is not included); the other rows use the same sources with different items.

	Kev-9B	Jev	JevK5	Jeeves v1	Jeeves v2
Test overall (out-of-domain and held-out)	0.822	0.857	0.802	0.884 (41%)	0.905 (39%)
MMLU-Pro and buried state	0.579	0.800	0.561	0.728 (86%)	0.780 (76%)
JevBench, public tiers	0.715*	0.866	0.861	0.913 (44%)	0.896 (29%)
JevBench hard	0.451*	0.730	0.748	0.820	0.793

* Kev-8B (Qwen3); no Kev-9B JevBench result is published.

Without thinking v2 scores 0.812 on the Jeeves test split, against 0.851 reasoning when unsure. On held-out questions from the 15 public datasets new in v2 it scores 0.853 and 0.900.

Serving on one H100 with --precision fp8, 325 dev questions:

setting	accuracy	median / p90 latency
full thinking	0.818	1.9 s / 8.1 s
nothink_threshold 0.9	0.822	1.2 s / 8.2 s
max_think 1024, nothink_threshold 0.9	0.818	1.2 s / 4.3 s
no thinking	0.806	22 ms / 46 ms

v2 was trained on chains of at most 1,024 tokens, so lower caps cut most of its chains.

Training

• SFT. LoRA r=16 on every projection plus the pointer head, 1 epoch (16,867 steps) on 453,531 questions from about 175 public sources. About 15% of examples carry a reasoning chain sampled from earlier rounds of the model, kept only if it closes within 1,024 tokens, does not loop and reaches the right answer.
• Calibration. Two softmax temperatures fitted on the dev set: 1.522 for direct answers and 5.420 for answers after reasoning. Every training chain reaches the right answer, so answers after reasoning come out overconfident and need the larger temperature.
• Drafters. Diffusion views of the frozen model, inspired by Orthrus, distilled by KL divergence on chains sampled from v1. They work with v2 (every draft is verified, so outputs are unchanged) but accept fewer tokens; v2 drafters will follow. At one question with v1 the block-4 drafter decodes 1.6× faster than graphed greedy decoding and the block-8 drafter 1.76×.

v1 was trained with SFT followed by CISPO; CISPO did not improve v2. The training code, dataset build script (python -m prep.prep, which rebuilds v2's training data byte for byte from pinned revisions) and the id lists of the mix are in the GitHub repository. The reasoning chains are not published yet.

Training data

Public sources only; no PostHog or customer data. Every new record was checked against the eval sets and dev for shared 8-word sequences. Each source stays under its own licence; the main groups are:

• Public datasets. AG News, Amazon reviews, ARC (CC BY-SA 4.0), Banking77, Bias in Bios, BoolQ (CC BY-SA 3.0), Civil Comments, CLINC150, CommonsenseQA, DBpedia (CC BY-SA 3.0), GoEmotions, IMDB, LIAR2, MASSIVE, MedMCQA, MedQA, Mind2Web, MNLI, Measuring Hate Speech, ProsocialDialog, QASC, SNLI (CC BY-SA 4.0), SST-5, TREC, WANLI, Yelp, HelpSteer2 and HelpSteer3, Anthropic HH-RLHF.
• New in v2. GSM8K, AQuA-RAT, RuleTaker, WinoGrande, HellaSwag, SWAG, Cosmos QA, OpenBookQA, the New Yorker caption contest, When2Call, ToolACE, Glaive function calling v2, RAGTruth, ESCI and Schema-Guided Dialogue (CC BY-SA 4.0).
• Third-party decision sets. Open-Jev v1 and v1.1 (CC0), the Kev suites, spark, typed-decisions, SemIf and decider's converted sets and teacher data. Their labels, and for several of them their texts, come from LLM teachers.
• Prompt injection. Gandalf, LLMail-Inject, the neuralchemy and S-Labs prompt-injection sets, InjecAgent and permissively licensed parts of ShieldLM's curation.
• Generated. Some sources are LLM-generated (for example Glaive, ToolACE, When2Call, Open-Jev and the teacher-labelled sets), and v1's synthetic policy data is generated from templates.

The full list of repositories, revisions and licences is in data/manifest.json in the GitHub repository.

Limitations

• Knowledge questions trail Jev (MMLU 0.845 vs 0.900, MMLU-Pro 0.770 vs 0.840).
• v2 trails v1 on JevBench when reasoning only when unsure (0.896 vs 0.913): it is overconfident without reasoning on some long-policy and multi-hop items, so it skips reasoning on them. With reasoning on every item the two are level (0.922 vs 0.926).
• Thinking is slow at the tail (8 s at p90 with full chains); use max_think and nothink_threshold when latency matters.
• Evaluated on English inputs only.

License

The weights are derived from Qwen3.5-9B and are released under its Apache-2.0 license (LICENSE). The Jeeves code is MIT licensed.

Citation

@software{waltz2026jeeves,
  author = {Waltz, Nicholas P.},
  title  = {Jeeves: Reasoning Improves Jev-like Decisions},
  year   = {2026},
  url    = {https://github.com/PostHog/jeeves},
  note   = {Qwen3.5-9B decision model that reasons when unsure, with a block-4 diffusion drafter}
}