JEV-27B for Core AI: typed decisions on your Mac
Run JEV-27B locally on an Apple silicon Mac. Give it a situation and a yes/no, multiple-choice or rating question; receive a probability for each allowed answer. This bundle brings AutoTrust AI's JEV-27B to Apple's Core AI framework with a Swift command line runner.
For: Swift developers building local typed-decision workflows. Not for: free-form chat, iOS, or high-risk decisions without independent validation. Requirements: Apple silicon, macOS 27, Xcode 27 with Metal Toolchain, Rust/Cargo and the Hugging Face CLI. GPU execution; measured on M4 Max with 128 GiB RAM. Minimum memory and sustained-use limits are not established; see Limitations.
Quick start
hf download okayuji/jev-27b-coreai --local-dir ./jev-27b-coreai
git clone https://github.com/oka-yuji/coreai-samples.git
cd coreai-samples
./Tools/Irodori/build-inputs.sh
swift build -c release --package-path Tools/Verify --product coreai-verify --force-resolved-versions
Tools/Verify/.build/release/coreai-verify typed-decide --bundle ../jev-27b-coreai/bundle --input ../jev-27b-coreai/examples/request.json --output ./decision-output.json
The output file must not already exist. Run one model job at a time; the companion repository documents its shared model-job lock in AGENTS.md. The bundle/ subfolder is intentional: the download cache and this README stay outside the strictly checked model files. Build prerequisites are described in the common setup.
Recorded output
The example request was written for this release and is not copied from an evaluation corpus:
The workshop opens at 09:00 and closes at 17:00 on weekdays.
Is the workshop open at 10:00 on a weekday?
The exact probabilities array from the recorded response, in false, true order, is:
[0.0008919818210415542, 0.9991080164909363]
The runner selected true. The request used 61 tokens and took 1.941008542 seconds after loading. This was a single first request without a warmup, recorded on 2026-11-04 on M4 Max 128 GiB, macOS 27.0 (26A428), Xcode 27 / Swift 6.4, Release, GPU preferred and one model operation at a time. Power and thermal conditions were not recorded for this release example. It is a usage check, not a latency benchmark; the controlled measurements below use warmed requests.
Key numbers
The following existing measurements were recorded on 2026-11-04 on M4 Max 128 GiB, macOS 27.0 (26A428), Xcode 27.0 (27A266a) / Swift 6.4, Release, GPU preferred, batch size and concurrency 1. AC power was supplied by a 140 W adapter reporting 94 W, with High Power mode. Every latency row started at nominal thermal pressure. Times are medians of three runs; each input length had one warmup. These measurements precede the release documentation changes and use the same asset, head, tokenizer and temperatures.
| Measurement | Result |
|---|---|
| Model bundle, excluding root documentation and examples | 13 files, 16,064,821,387 bytes (14.96 GiB) |
| One decision, 100 tokens | 0.615 s |
| One decision, 500 tokens | 2.208 s |
| One decision, 2,000 tokens | 8.978 s |
| One decision, 8,192 tokens | 37.139 s |
| Fresh-specialization load | 28.224 s |
| Load from the specialization cache in a new process | 9.505 s |
| Process lifetime peak RSS | 32.33 GB |
| Process lifetime peak physical footprint | 21.34 GB; not a bound |
| Largest sampled RSS during an 8,192-token request | 17.30 GB |
Decision time includes rendering, tokenization, backbone inference and the head, but excludes loading and JSON file I/O. Load times include file verification, tokenizer, head and backbone. RSS and footprint are different measures and must not be added. During repeated requests of alternating lengths, footprint kept increasing; see Limitations. One long measured request encountered thermal pressure after it started; the median rows ran at nominal pressure. Three runs do not establish p95. Earlier battery and 70 W measurements remain in the full performance record.
No controlled speed comparison with the original model or Core ML is claimed. Sources: performance-ac94.json and measurement conditions. The evaluated bundle before documentation changes was 16,064,815,496 bytes; that historical measurement is retained in the companion record.
What is included
The System 1 adapter is merged into the text backbone, then quantized to uint4 with the official qwen3_5_cfg recipe (block size 32, minval). Gated delta updates run in float32 to preserve the checked numeric fidelity. A single prefill produces the final hidden state; the unchanged 24-row float32 head returns the typed scores. Per-kind temperatures were refitted on the Core AI outputs. The single-forward decision architecture and the original distillation are upstream JEV-27B features.
| Path | Purpose |
|---|---|
bundle/*.aimodel/ |
Converted backbone, its identity and asset metadata |
bundle/head.safetensors |
Original float32 decision head |
bundle/tokenizer/ |
Byte copies of the pinned upstream tokenizer files |
bundle/decision.json |
Typed-decision contract, temperature values, provenance and all model-file sizes/SHA-256 values |
bundle/calibration-refit.json |
Temperature-refit settings, aggregate statistics and excluded-row identifier digests |
bundle/SHA256SUMS |
Checksums for every other file in bundle/ |
LICENSE, NOTICE.md |
License, attribution and file-level modification notices; also retained inside the bundle |
examples/ |
An authored request and the exact recorded runner output |
The evaluated decision manifest had SHA-256 a2b186a87a64d9e93a56f185a5ebc2a0613d00da7440eeb8c8c2229b64f6adf3. The release updates notices, metadata and checksum manifests. The backbone asset, head and tokenizer remain byte-identical. The public calibration record replaces all excluded-row identifiers with their UTF-8 SHA-256 digests in the original order. Temperatures, fit conditions, fixture/input hashes and numerical statistics are unchanged; this is a publication representation of the existing calibration, not a new fit. Its source_schema and source_record_sha256 identify the original record, while decision.json uses the hash and calibration ID of the public representation. Check the release bundle with shasum -a 256 -c SHA256SUMS from inside bundle/, or use the companion recipe's decision.py --check-files entry point.
Input contract
The Swift runner accepts JSON rows with id, state, question, kind and options. It renders the upstream bare-v1 template as one string, without adding special tokens:
[kind] {kind}
[state] {state}
[question] {question}
[options]
{options}
[decision]:
There is no newline after [decision]:. noul requires false, true; score requires 0 through 5; choice accepts 2 to 16 distinct single-line options, rendered as A) ... through P) .... Input limits apply to the complete rendered string: 8,192 tokens. Unsupported options and longer inputs are rejected before inference. The lowest option index wins an exact tie. See the Swift API and file contract.
Verification
The companion repository preserves the original conditions, failures and missing measurements in the validation records. Fidelity and quality were measured on M4 Max 128 GiB / macOS 27.0 (26A428), GPU preferred, one model operation at a time, on 2026-11-03 to 2026-11-04. Power and thermal conditions were not recorded for these older quality runs.
- On F, 301 accepted inputs spanning up to 8,192 tokens, plus one rejected over-limit input, compare the Core AI asset with PyTorch float32 using identical quantized weights. At temperature 1, maximum probability difference is 0.0065415, with no top-option change. Every populated length band meets the fixed maximum 0.03 and mean 0.01 limits.
- On Q, 900 corpus rows, quantization and conversion retain the recorded adaptation gains. Choice agreement with the teacher top (n = 300) is 86.67%, versus 87.33% before quantization. The difference is -0.67 percentage points, with a 95% interval of [-3.00, +1.67] that crosses the -2-point threshold. The existing pass is based on the point estimate, not a statistical proof of non-inferiority. noul Brier changes by +0.00024 and normalized score MAE by +0.00051.
- Temperatures are noul 1.0056266, choice 1.0053791 and score 1.010558. They were fitted on Cal and selected on H' after removing shared states; NLL and multiclass Brier are no worse than at the upstream card temperatures. These are measures of agreement with the teacher distributions, not calibration against real-world correctness.
- Japanese JGLUE results cover 200 rows per task and have no pass line. They are descriptive results, not a certification of Japanese deployment quality. The prompt variants and selection are recorded separately.
Sources: fidelity.json, quality.json and Japanese evaluation.
Limitations
- Repeated-use memory growth is unresolved. In 16 requests alternating 100, 500, 2,000 and 8,192 tokens, footprint during the 8,192-token rows grew from about 8.2 to 12.5, 16.6 and 21.2 GB. It increased by about 4.1 to 4.5 GB per round and did not level off. Long-running deployment and lower-memory Macs have not been validated. The stated peak is not a memory requirement or upper bound.
- The adapter merge has a known numeric gap. The unmerged adapter and merged checkpoint differ by up to 0.038 in probability, above the original 0.01 merge-check limit. This was retained as a limited exploration with the failed check recorded; the threshold was not relaxed. The later fidelity pass compares Core AI to the same merged and quantized weights and does not erase this gap.
- Scope: English-centered text decisions through the macOS command line tool. No System 2 generation, no choices above 16, no iOS validation and no native-app integration for this model. The minimum supported Mac memory has not been established; the recorded machine has 128 GiB.
- Long inputs: fidelity was checked up to 8,192 tokens, but quality and calibration above about 700 tokens were not measured. Long-input numerical agreement does not establish useful decision quality.
- Probabilities: teacher-distribution agreement is not real-world correctness. Do not use the model as the sole decision maker for medical, legal, financial or other high-risk decisions.
- Cancellation and lifetime: the recorded cancellation test used a backbone wrapper that does not forward cancellation and completed after its active run returned. It does not establish cancellation of Core AI's run itself. Release/reload within one process and document-prefix reuse are unverified.
- Version: this is a conversion of the pinned
autotrust/JEV-27Bsource, notAutoTrust-JEV-27B-v1.1, whose inference contract differs.
License and attribution
The model weights inherit Apache License 2.0 from Qwen3.8-27B and JEV-27B; see LICENSE and NOTICE.md for the original copyright notice, pinned revisions, corpus credit and file-level changes. The companion Swift code is MIT licensed in its own repository.
This is an unofficial Core AI conversion by okayuji and is not endorsed by or affiliated with AutoTrust AI or TypeSafe. The upstream System 1 adapter and head were trained on SargeDev's jev-distill-corpus-v3, largely on output distributions of the TypeSafe Jev 1.13 teacher. That corpus also supplied this conversion's evaluation and calibration sets. Neither its texts nor JGLUE texts are distributed here. Names identify their respective works and owners; no endorsement is implied.
For provenance or license concerns, use model discussions or the companion issue tracker. A substantiated concern will be investigated, with the affected distribution temporarily suspended while its provenance and permission to distribute are reassessed.