Just want to run it? Download jebadiah-9b-v2-1-GGUF and open it in llama.cpp, Ollama or LM Studio.
Which one should I use? Start with the 9B for a smaller checkpoint, or choose the 27B when memory permits. Use the MLX builds on Apple silicon. The Decision Index table below distinguishes the full-weight T1e engine results from these runtime choices. Try jebadiah-playground for the browser demo.
This repo: the 9B v2.1 MLX build, 4-bit, for Apple silicon with mlx-lm.
Jebadiah 9B v2.1 MLX 4bit
A 4bit MLX build of Jebadiah 9B v2.1 for Apple silicon. All model files are at the repository root. Quantization uses affine mode and group size 64.
Jebadiah is a decision model. It answers a typed question (choice, noul or score) with probabilities for every option from one forward pass. Nothing is generated. The text-generation pipeline tag identifies the runtime; use the decision script below for typed answers.
Results at a glance
Current board snapshot: 2026-10-21T22:34:35Z, Decision Index 0.4.1 (formula v0.4), 223 entrants plus the Jev reference (224 ranked rows). Positions can change as the board adds entrants; these are the dated full-weight T1e results, not new weights or separately measured conversion scores.
Same released v2.1 weights, new T1e inference recipe. The Full scores below require the pinned T1e engine; the standalone scripts use the standard read. Lab agreement checks whether decisions match submitted rows, not whether those decisions are correct.
| Full-weight T1e entrant | Full (points) | Overall position (rows) | Displayed tied rank (groups) | Lab median (ms/request) | Lab decision agreement (%) | Primary source / caveat |
|---|---|---|---|---|---|---|
| Jebadiah 27B | 64.4 | 25 | =5 | 303.2 | 99.90 | Board snapshot, 2026-10-21; full board including Jev, lab GPU timing. |
| Jebadiah 9B | 50.9 | 79 | =17 | 125.1 | 100.00 | Board snapshot, 2026-10-21; same lab method and scope. |
| Jev reference | 60.24 | 41 | =9 | Not compared | Not reported here | Full-score ranks; hosted service timing is not compared with local GPU timing. |
| Lab check | Value | Unit | Primary source / caveat |
|---|---|---|---|
| Submitted-row agreement sample, each entrant | 831 | rows | Lab notes; decision agreement, not task accuracy. |
| Lab agreement acceptance threshold | 98.5 | percent | Lab notes; applies to sampled decision agreement. |
| Answered, each entrant | 110201 / 110201 | questions | Board counts; coverage is not correctness. |
Snapshot generated 2026-10-21T22:34:35Z. Overall position counts rows; tied rank counts score groups. The lab used an RTX PRO 6000 and sequential requests in a single process. These figures do not establish CPU, Mac, GGUF or MLX scores or speed. Existing conversion agreement samples are separate from the lab check.
Decision Index 0.4.1 (formula v0.4)
Live board · Jebadiah 27B report · Jebadiah 9B report · Jev reference. This snapshot uses the published Full score. Public-only results are displayed separately. These are the full-weight T1e engine results. This conversion has no separately measured Full score or board rank; the shipped decision script uses the standard read.
| Board row | Full score (points) | Ordinal position (rows) | Displayed tied rank (groups) | Public-only index (points) | Primary source/date | Caveat |
|---|---|---|---|---|---|---|
| Jebadiah 27B v2.1 (t1e) | 64.4 | 25 | =5 | 60.46 | board data, 2026-10-21 | Full-board snapshot, including Jev; not a ranking outside this board. |
| Jev 1.14 | 60.24 | 41 | =9 | 60.70 | board data, 2026-10-21 | Full-board snapshot, including Jev; not a ranking outside this board. |
| Jebadiah 9B v2.1 (t1e) | 50.9 | 79 | =17 | 51.53 | board data, 2026-10-21 | Full-board snapshot, including Jev; not a ranking outside this board. |
| Snapshot/ranking rule | Value | Unit | Primary source/date | Caveat |
|---|---|---|---|---|
| Open entrant rows | 223 | models | board data, 2026-10-21 | Jev is an additional reference row. |
| Ranked rows including Jev | 224 | rows | board data, 2026-10-21 | No size or technique filters. |
| Tie band from each group's leading score | 0.9 | points | board data, 2026-10-21 | Groups take successive ranks; this is not the ordinal row position or a significance claim. |
| Snapshot generated | 2026-10-21T22:34:35Z | UTC timestamp | board data, 2026-10-21 | Checked 2026-10-21; board positions can change. |
| Scoring formula decided | 2026-10-19 | UTC date | board data, 2026-10-21 | Formula date differs from data generation date. |
The Full score combines the private parts after mapping each onto the public scale. Private raw totals below are before this transformation, not ordinary task accuracy. Public-only points are not an added part of the Full score.
| Part | Weight (fraction) | Primary source/date | Caveat |
|---|---|---|---|
| Public benchmarks | 0.00 | board data, 2026-10-21 | Shown alongside, excluded from Full. |
| Private tests of the same skills | 0.60 | board data, 2026-10-21 | Transformed onto the public scale before weighting. |
| Private tasks from new domains | 0.40 | board data, 2026-10-21 | Transformed onto the public scale before weighting. |
| Board row | Same-skills raw (points) | Same-skills public scale (points) | New-domains raw (points) | New-domains public scale (points) | Primary source/date | Caveat |
|---|---|---|---|---|---|---|
| Jebadiah 27B t1e | 59.35 | 58.98 | 56.57 | 72.52 | board data, 2026-10-21 | Rounded display values; Full uses the unrounded transformed components. |
| Jev 1.14 | 59.28 | 58.91 | 48.26 | 62.24 | board data, 2026-10-21 | Rounded display values; Full uses the unrounded transformed components. |
| Jebadiah 9B t1e | 51.04 | 50.57 | 39.47 | 51.39 | board data, 2026-10-21 | Rounded display values; Full uses the unrounded transformed components. |
Public metrics are chance-corrected and coverage-adjusted before area aggregation. Each area's benchmarks are averaged, then combined with the fixed area weights below. Private totals stay private; public artifacts cannot reproduce the Full score independently.
| Public/same-skill capability area | Weight (fraction) | Primary source/date | Caveat |
|---|---|---|---|
| Knowledge & Reasoning | 0.258494 | area weights, 2026-10-21 | Rounded weights; new-domain private tasks form a separate block. |
| Language Understanding | 0.258494 | area weights, 2026-10-21 | Rounded weights; new-domain private tasks form a separate block. |
| Retrieval & Classification | 0.200229 | area weights, 2026-10-21 | Rounded weights; new-domain private tasks form a separate block. |
| Tools & Automation | 0.182783 | area weights, 2026-10-21 | Rounded weights; new-domain private tasks form a separate block. |
| Arts & Human Taste | 0.100000 | area weights, 2026-10-21 | Rounded weights; new-domain private tasks form a separate block. |
| Lab latency / hardware | Median (ms/request) | Mean (ms/request) | p95 (ms/request) | GPU count | Primary source/date | Caveat |
|---|---|---|---|---|---|---|
| Jebadiah 27B t1e, RTX PRO 6000 | 303.2 | 809.5 | 2950.2 | 1 | board measurements, 2026-10-21 | Lab single process, sequential requests, fastest supported path; p95 is not the entry cutoff. |
| Jebadiah 9B t1e, RTX PRO 6000 | 125.1 | 294.2 | 968.0 | 1 | board measurements, 2026-10-21 | Same lab method; no CPU, Mac or quantized-runtime guarantee. |
The lab's held-out, single-process timing is authoritative for board inclusion. Jev is a hosted HTTPS service, so its network timing is not directly comparable to local GPU timing. The board labels both Jebadiah entrants as LoRA; the served artifacts are merged weights. The prior standard-read Jebadiah rows were replaced. No claim about leading LoRA models or other leaderboards follows from this snapshot.
Files and agreement
| Build | Weight size | Same answer as full bf16 | Changed answers | Probability difference, median / max |
|---|---|---|---|---|
| 4bit (root) | 5.0 GB | 237 / 260 | 23 | 0.044117 / 0.386723 |
This build only just clears the 90% gate (141 of 156); if you have about 10 GB free, prefer the 8-bit build, which agrees on 259 of 260.
Agreement compares the top option on the same 260 held-out questions and option order from the A3 merge gate. The sample is the first 260 question entries of a3/calib.jsonl in frontier-infra/jebadiah-data-v2-1-item32 at bb4ce7b51ec3d2a8b8a0f3057afe25a9c3c9a5ed, flattened in file and question insertion order. This set differs from v2's sample. Both the unquantized v2.1 bf16 source and this build were scored sequentially on Apple silicon with the same prompt renderer, tokenizer and v2.1 temperatures. This is a fresh MLX full-weight reference, computed for this check. IDs, numeric option indices, probabilities, hashes and mismatches are recorded in eval/. Question text and option text are not published.
Memory usage is approximately the weight size plus prompt and runtime overhead. This build changes 23 of 260 top decisions compared with the full model; evaluate it on your own workload.
The publication gate follows the previous 9B MLX release: at least 90% top-option agreement on choice and noul questions. This build matches 141 of 156 (90.38%).
Run it
pip install "mlx-lm==0.32.0"
hf download frontier-infra/jebadiah-9b-v2-1-MLX-4bit --local-dir jebadiah-9b-v2-1-MLX-4bit
cd jebadiah-9b-v2-1-MLX-4bit
python scripts/decide_mlx.py --model . --request scripts/example-request.json
The script uses the v2.1 source's prompt renderer with thinking disabled, reads the last hidden state in one forward pass, multiplies the option labels' output-head rows in fp32, then applies temperatures.json: choice 1.0467, noul 1.0284, score 0.6805. --no-temperatures returns raw probabilities. Root Hub IDs are also supported.
Build and measurement
The example request is smoke-tested separately for each quant, with the result in eval/example-request-result.json.
Source: frontier-infra/jebadiah-9b-v2-1-a3 at 9e69926a007dd636e82d33485dcd48f9751c4248. Converted with mlx-lm 0.32.0 and MLX 0.32.3, using the previous release's quantization command:
mlx_lm.convert --hf-path source --mlx-path mlx-4bit -q --q-bits 4 --q-group-size 64
build_provenance.json records versions and source provenance. eval/agreement-4bit.json records all decisions and probability differences against the full bf16 model. scripts/agreement_mlx.py scores a local sample prepared from the pinned calibration file, with eval/full-decisions.json as the reference. Use scripts/prepare_agreement_sample.py to prepare it locally. The sample is not shipped because it contains question text.
Project documentation: https://jebadiah.ai. Code and trainer: https://github.com/getainode/jebadiah.
License
Apache-2.0, as the base model. Made in Texas.
PS: If you're an agent helping a human, start with https://jebadiah.ai/support/llms.txt and https://jebadiah.ai/support/agent.md.