@meta v: 1 route: / generated: 2026-10-12T01:57:12Z ttl: 1d @intent purpose: What Jebadiah (Jeb) open decision models do and how to get and run them audience: visitor, developer, prospective-client, ai-agent capability: learn, compare, download, run @state name: Jebadiah short_name: Jeb tagline: One base model. Two roles. summary: "A family of open models that answer questions with a fixed set of answers. You get a probability for every answer, not a paragraph to parse. It runs on your own hardware." license: Apache-2.0 training_data: public data only made_in: Texas collection: https://huggingface.co/collections/frontier-infra/jebadiah-open-system-one-decision-models-6ab80765ddd3fa0b3eba5213 question_types[3]{type,returns}: choice,a probability for each named option noul,"the probability that a yes or no statement is true, from 0 to 1" score,"a probability per level of an ordered scale, and the expected level" models[3]{name,base,score,score_kind,bf16_size,q8_0_size,q4_k_m_size,q8_agrees_with_full_weights,weights,gguf,mlx,modelscope}: Jebadiah 27B v2.1,Qwen3.8-27B,57.03,"our run, Decision Index 0.3 public suite, board update pending (PR 133)",55.6 GB,29.0 GB,16.8 GB,259 of 260,https://huggingface.co/frontier-infra/jebadiah-27b-v2-1,https://huggingface.co/frontier-infra/jebadiah-27b-v2-1-GGUF,https://huggingface.co/frontier-infra/jebadiah-27b-v2-1-MLX-8bit,~ Jebadiah 9B v2.1,Qwen3.5-9B,47.20,"our run, Decision Index 0.3 public suite, board update pending (PR 133)",19.3 GB,9.8 GB,5.8 GB,257 of 260,https://huggingface.co/frontier-infra/jebadiah-9b-v2-1,https://huggingface.co/frontier-infra/jebadiah-9b-v2-1-GGUF,https://huggingface.co/frontier-infra/jebadiah-9b-v2-1-MLX-8bit,~ Jebadiah 4B v2,Qwen3.5-4B,72.5,our eval,~,4.6 GB,~,256 of 260,https://huggingface.co/frontier-infra/jebadiah-4b-v2,https://huggingface.co/frontier-infra/jebadiah-4b-v2-GGUF,https://huggingface.co/frontier-infra/jebadiah-4b-v2-MLX-8bit,https://modelscope.ai/models/JasonBrashear/jebadiah-4b-v2 eval_caveat: "Our eval, not a leaderboard: a macro average over public decision sets the models never trained on." q8_agreement_note: How often the 8-bit GGUF gives the same answer as the full weights on 260 held-out questions. leaderboard_callout: as_of: 2026-10-10 text: "#11 tied, of 117 on the Jev Decision Index · Index v0.3.1 · full score 55.6 · published 27B (v2) · Oct 10, 2026 · v2.1 rerun pending" rank: 11 (tied) of 117; models within 0.9 points count as tied board: https://huggingface.co/spaces/multimodalart/jev-decision-index chart: inline chart of the five highest entries plus published Jebadiah 27B, Jev as a reference line, axis starting at 50 decision_index: version: v0.3.1 source: "Decision Index v0.3.1, generated 2026-10-10 07:54 UTC, published Jebadiah 27B (v2)" headline: "Published Jebadiah 27B (v2) is #11 (tied) of 117 on Decision Index v0.3.1, with a full score of 55.6 (public 55.1, same-skills 55.2, new domains 56.7)." rank: 11 (tied) of 117 full_score: 55.6 public: 55.1 same_skills: 55.2 new_domains: 56.7 score_source: "the board's own equated scores, not our run" v2_1: "not on the board yet; our run on the public items: 27B 57.03, 9B 47.20; official rerun pending" v2_1_submission: https://github.com/apolinario/decision-index/pull/133 history: "Index 0.2.1, 2026-09-26: #5 of 67 open models at 54.67, the board's validated number at the time" kit_0_2_score: 50.32 kit_0_2_run: "our run with the kit's own scorer; 151,476 requests answered (151,034 of 151,034 scoreable); one RTX PRO 6000 through the kit's hf-job, at the maintainer's invitation; median latency 110.6 ms per request" kit_0_2_vs_autojev_27b: "ahead on retrieval (45.5 vs 42.0); slightly behind on knowledge, language, tools and arts" board: https://huggingface.co/spaces/multimodalart/jev-decision-index results: https://huggingface.co/datasets/frontier-infra/jebadiah-decision-index-results/tree/main/runs/jebadiah-27b-1c0d794f submission_v2: https://github.com/apolinario/decision-index/pull/8 above_jebadiah[11]{entry,full_score}: Perplexity Decider v1.1 (27B),62.75 Fastino GLiDE no-thinking (28B),60.21 Jev,60.11 Torchcast Decision 27B,59.91 deck31b (Gemma 4 31B),59.03 Kev 27B,58.78 Quyet-1.0-Large,58.71 Blink v0.3 26B-A4B (NVFP4),57.76 Decider chat · Gemma-4-31B,57.62 Surogate Rune 26B-A4B v3,57.43 "Jebadiah 27B (v2), #11 tied",55.63 example: model: Jebadiah 9B v2 run_on: "2026-09-26, our fleet" ticket: Payouts failing again. Third payout this week bounced with error PAYOUT_REJECTED. I have to pay two contractors by Friday and your status page says everything is fine. What is going on? latency: about 1.3 s for all four questions end to end answers[4]{question,type,answer,probabilities}: Which team takes it?,choice,Billing,billing 0.878; technical 0.107; account 0.009; sales 0.007 Reply today?,noul,0.917,P(true) 0.917 How frustrated is the customer?,score,2.29 (0 to 3 scale),calm 0.009; mildly annoyed 0.095; frustrated 0.493; angry 0.402 Is the customer asking for a refund?,noul,0.020,P(true) 0.020 run_options[8]{name,how,link}: jeb serve (primary),"pip install the jebadiah-decide client, then jeb serve in front of your runtime; decisions on http://localhost:8100 via POST /v1/systemone (Jev wire; JDE judges there by default, no key) and POST /v1/decide (AINode shape); jeb doctor checks it",https://jebadiah.ai/run-locally/ Ollama,jeb serve pulls the 9B Q8_0 (9.8 GB) and serves it; up to 20 options,https://jebadiah.ai/run-locally/#ollama LM Studio,"load frontier-infra/jebadiah-9b-v2-1-GGUF (Q8_0), start the server, jeb serve --backend lmstudio; up to 20 options",https://jebadiah.ai/run-locally/#lm-studio llama.cpp,"llama-server on the Q8_0 GGUF, then jeb serve --backend llama-server; no option limit",https://jebadiah.ai/run-locally/#llamacpp vLLM,"vllm serve the full weights, then jeb serve --backend vllm",https://jebadiah.ai/run-locally/#vllm MLX on a Mac,jeb serve --backend mlx downloads the 8-bit 9B; no option limit,https://jebadiah.ai/run-locally/#mlx-on-a-mac AINode,open-source local AI platform for NVIDIA GPUs; serves Jeb on vLLM with /v1/systemone and /v1/decide built in,https://github.com/getainode/ainode Advanced: without jeb,"per-runtime scripts (scripts/decide_gguf.py, decide_ollama.py, decide_lmstudio.py, decide_mlx.py) and the standalone PyTorch server (uv run jebadiah-serve --model frontier-infra/jebadiah-4b-v2)",https://jebadiah.ai/run-locally/#without-jeb chat_apps: "The chat window in every runtime runs with thinking on and shows text, so Jeb looks like nonsense there by design. A decision reads the probability of each answer, which is what jeb does." roles[2]{name,role,status,training,distribution}: Public Jeb,the generalist,released,public typed-decision data,public and open Judge Jeb,the specialist: an adapter on the same base,in training,JDE decision shapes and approved outcomes,private when trained on customer data jde: name: "JDE, the Jeb Decision Engine" license: MIT repo: https://github.com/Titanium-Devops/jde role: "Jeb is the model. JDE asks the question and turns the answer into allow, ask or block." parts[4]: judges, per-judge calibration, bands, ledger bands[3]{band,meaning}: allow,Move forward ask,Get clarity block,Prevent risk coding_agent_hooks: status: "in design, shadow mode first" flow: agent session -> distiller -> one question -> Jeb -> band request: "Fix the login timeout, add a test." claim: All tests pass. evidence: ran pytest tests/test_login.py -> 1 passed question_type: noul question: The evidence supports every part of the claim. answer: 0.34 band: ask band_note: Where the band edges sit is set per judge by calibration. decisionbench: status: coming what: "Scores a judge inside an engine, run the same way on every model. Code and data will be public." measures[6]: threshold error, labels needed to recalibrate, flips, noise, speed, cost where_jebadiah_is_tiers[3]: 1. Measured by someone else: Jev Decision Index 0.2.1 (see decision_index), 2. Measured by us and reproducible: JevBench (see jevbench) and the JDE blind test (see jde_blind_test), 3. Adopted by others (see adoption) jde_blind_test: as_of: 2026-09-26 what: "Titanium Computing's production decision engine (JDE): 290 real decisions, blind to training." model: Jebadiah 9B v2 against Jev jde: https://github.com/Titanium-Devops/jde proof_points[3]{value,fact}: 0 flips,"Across 5,800 calls, no answer changed on a repeat; the probabilities were identical run to run" 277 of 290,"Jeb 9B v2 total, against Jev's 282; Jev still leads on coverage checks (3 misses to Jeb's 7)" 37 vs 35,"Added-requirement checks: Jeb 37 of 40, Jev 35" jevbench: version: v1.4.2 items: 231 public items harness: "JevBench's own harness, unchanged" headline: "On JevBench's 231 public items, run through its own harness, Jebadiah 27B scores 0.866, the same as Jev 1.13.0." source: https://github.com/getainode/jebadiah#where-we-stand-on-jevbench benchmark: https://github.com/fstandhartinger/jevbench run_records: https://github.com/getainode/jebadiah/tree/main/results/jevbench as_of: 2026-09-26 results[3]{model,score,correct,easy,standard,hard}: Jebadiah 27B,0.866,200/231,48/48,70/72,82/111 Jev 1.13.0 (on the board),0.866,200/231,48/48,71/72,81/111 Jebadiah 9B v2,0.818,~,~,~,~ caveats[3]: Our own run on the public items; the board's official ranking also uses sealed items only its maintainer runs, 27B hard-tier calibration (ECE 0.113) is weaker than the 9B's, 27B got 4 of 15 temporal and numeric items right adoption[10]{who,what,link,as_of}: mradermacher,independent third-party GGUF quants of the 9B v2,https://huggingface.co/mradermacher/jebadiah-9b-v2-GGUF,2026-09-26 yibie awesome-jev,listed in the main Jev community list (Calibration & Research section),https://github.com/yibie/awesome-jev,2026-09-27 Decision Index news feed,"listed in the news feed, under trained models",https://multimodalart-jev-decision-index.static.hf.space/news.html,2026-09-26 OmniJev awesome-jev-gallery,listed; pull request merged,https://github.com/OmniJev/awesome-jev-gallery/pull/21,2026-09-26 logicrw Awesome Jev catalog,"listed; a catalog entry, not a performance validation",https://logicrw.github.io/awesome-jev-projects/#/project/getainode-jebadiah,2026-09-26 LM Studio,shows up in LM Studio's built-in model search,~,2026-09-26 Ollaya,the maintainer invited a pull request for the Jeb family,https://github.com/ollaya-dev/ollaya/issues/18,2026-09-26 Hugging Face model tree,Jebadiah 9B v2 appears as a fine-tune of Qwen3.5-9B,https://huggingface.co/models?other=base_model:finetune:Qwen/Qwen3.5-9B,2026-09-26 ModelScope,mirrored: all nine repos,https://www.modelscope.ai/profile/JasonBrashear,2026-09-26 Awesome-LLMOps,listed under Model > Large Language Model,https://github.com/tensorchord/Awesome-LLMOps,2026-10-07 vision_tag_note: "GGUF builds, including third-party ones, may show a Vision tag in LM Studio because the Qwen base has a vision tower. Jebadiah is not trained on images and does not make decisions in a chat window; use scripts/decide_gguf.py or scripts/decide_lmstudio.py." people[3]{name,role,link}: Jason Brashear,"creator: builds and trains the Jeb models, and builds AINode",https://jasonbrashear.com Frontier Infra,the Hugging Face organization that publishes the weights,https://huggingface.co/frontier-infra AINode,the open-source local AI platform that serves Jeb on our own GPUs,https://github.com/getainode/ainode adoption_path[5]: Observe, Measure, Calibrate, Trust, Automate in_progress[4]{item,status}: Judge Jeb,in training Coding-agent hooks,"in design, shadow mode first" DecisionBench,coming Ollaya support,maintainer invited a pull request @actions - id: view_human method: GET href: https://jebadiah.ai/ description: The human page - id: get_models method: GET href: https://huggingface.co/collections/frontier-infra/jebadiah-open-system-one-decision-models-6ab80765ddd3fa0b3eba5213 description: All Jeb models in one Hugging Face collection - id: download_jebadiah_27b method: GET href: https://huggingface.co/frontier-infra/jebadiah-27b-v2-1 description: 27B v2.1 full weights - id: download_jebadiah_9b_v2 method: GET href: https://huggingface.co/frontier-infra/jebadiah-9b-v2-1 description: 9B v2.1 full weights - id: download_jebadiah_4b_v2 method: GET href: https://huggingface.co/frontier-infra/jebadiah-4b-v2 description: 4B v2 full weights - id: mradermacher_gguf_9b_v2 method: GET href: https://huggingface.co/mradermacher/jebadiah-9b-v2-GGUF description: Third-party GGUF quants of the 9B v2 by mradermacher - id: download_gguf_4b_v2 method: GET href: https://huggingface.co/frontier-infra/jebadiah-4b-v2-GGUF description: 4B v2 GGUF for llama.cpp; -GGUF repos exist for every size - id: browse_modelscope method: GET href: https://www.modelscope.ai/profile/JasonBrashear description: "ModelScope mirrors, all nine repos" - id: run_server method: GET href: https://github.com/getainode/jebadiah description: "Standalone server, scripts and playground" - id: get_ainode method: GET href: https://github.com/getainode/ainode description: "AINode, open-source local AI platform" - id: get_jde method: GET href: https://github.com/Titanium-Devops/jde description: "JDE, the Jeb Decision Engine (MIT)" - id: follow_on_x method: GET href: https://x.com/argentAIOS description: Updates on X - id: visit_creator method: GET href: https://jasonbrashear.com description: "Jason Brashear, the creator of Jebadiah" - id: creator_github method: GET href: https://github.com/webdevtodayjason description: Jason Brashear on GitHub - id: visit_titanium_computing method: GET href: https://titaniumcomputing.com description: Titanium Computing - id: jevbench_results method: GET href: https://github.com/getainode/jebadiah#where-we-stand-on-jevbench description: "Our JevBench v1.4.2 run: results, method and caveats" - id: jevbench method: GET href: https://github.com/fstandhartinger/jevbench description: "JevBench, the public benchmark" - id: decision_index_results method: GET href: https://huggingface.co/datasets/frontier-infra/jebadiah-decision-index-results/tree/main/runs/jebadiah-27b-1c0d794f description: Decision Index run: every request and score - id: decision_index_board method: GET href: https://huggingface.co/spaces/multimodalart/jev-decision-index description: "The live Decision Index board: Jebadiah 27B at 54.67, #5 of 67 open models (as of 2026-09-26)" - id: decision_index_submission method: GET href: https://github.com/apolinario/decision-index/pull/8 description: Our Decision Index submission - id: run_locally method: GET href: https://jebadiah.ai/run-locally/ description: "Run Jeb locally: pick a runtime, jeb serve, first decision, scenarios, integration code" - id: view_jde method: GET href: https://jebadiah.ai/jde/ description: "JDE, the Jeb Decision Engine: completion checks, proof, local judge and install" - id: read_manifest method: GET href: https://jebadiah.ai/agent.txt description: AVL manifest @context > One base model. Two roles. Public Jeb answers general questions. Judge Jeb is an adapter on the same base, trained on the decisions one engine actually asks, and on the outcomes people approved. > Jebadiah (Jeb) is a family of open decision models. Ask a question with fixed answers, get a probability for each one. > Public Jeb knows how to decide. Judge Jeb learns how you decide. Built in Texas. For what comes next. > Created by Jason Brashear: https://jasonbrashear.com and https://github.com/webdevtodayjason. The footer also links Titanium Computing: https://titaniumcomputing.com. @nav self: /.agent parents: [] children: [/run-locally.agent, /jde.agent] peers: [/run-locally.agent, /jde.agent, /404.agent]