Shingi 27B — 審議
Shingi 27B is a local decision model. You give it context (text, and optionally images), a question and the possible answers. It returns a choice, calibrated probabilities or an ordinal score. It reads the logit of every candidate answer directly and generates no text.
The model is a single 7.2 GB ternary GGUF file (PQ2_0). It is Bonsai 2 27B with retrained block scales: the ternary weights are unchanged, only their scales were trained. It needs no adapter.
Run it
On Linux with an NVIDIA GPU and the CUDA toolkit, or on a Mac with Apple Silicon:
curl -fsSL https://raw.githubusercontent.com/kortexa-ai/shingi-27b/main/run.sh | bash
This builds the pinned Prism runtime,
downloads this model into your Hugging Face cache and serves the API on
http://127.0.0.1:8765. Use --port to change the port. The code, API examples
and options are in kortexa-ai/shingi-27b.
Requirements
| Resource | Need |
|---|---|
| GPU | NVIDIA with at least 20 GB: RTX 4090, RTX PRO 6000 or DGX Spark (tested). The model uses about 8 GiB at the 16K context, about 9 GiB with image input. |
| System RAM | about 1 GB for the server process, plus page cache for the 7.2 GB file |
| Disk | 7.2 GB for the weights, plus the runtime build |
| Mac | Apple Silicon with 24 GB or more of unified memory (16 GB minimum), using Metal |
| Software | Linux (x86-64 or aarch64) with the CUDA toolkit, or macOS with the Xcode command line tools; CMake and a C++17 compiler |
Speed
Median latency per decision, one request at a time, loading excluded:
| Request | RTX PRO 6000 | RTX 4090 | DGX Spark |
|---|---|---|---|
| Short yes/no (~120 tokens) | 66 ms | 85 ms | 148 ms |
| Short choice, 3 options | 73 ms | 95 ms | 181 ms |
| Choice, 10 options (~1,300 tokens) | 246 ms | 306 ms | 802 ms |
| Long choice, 3 options (~3,000 tokens) | 881 ms | 1,045 ms | 2,984 ms |
| GPU memory in use | 8.3 GB | 8.2 GB | 7.9 GB |
With image input on the RTX 4090, one image and one question take about 0.66 s (median) and the model uses about 9.2 GB.
Questions in one request share a single pass over the images and the state, so extra questions are cheap: on an RTX PRO 6000, eight questions about one image take about 1.8 s instead of 8.7 s.
Scores
External suites, canonical choice order, 16K context:
| Suite | Items | Accuracy |
|---|---|---|
| JevBench public | 231 | 85.7% |
| JevBench hard | 111 | 72.1% |
| DecisionBench | 586 | 75.3% |
| DecisionBench hard | 293 | 65.2% |
| This/That | 7,305 | 67.9% |
These suites also informed the choice of training data, so treat them as development results rather than a clean held-out test.
More external suites. None of their items are in the training data, but Mind2Web test shares websites with the Mind2Web train split we train on:
| Suite | Metric | Bonsai 2 27B (base) | Shingi 27B | Jev 1.13.0 (reported) |
|---|---|---|---|---|
| PubMedQA (1,000) | accuracy | 72.1% | 78.8% | 78.7% |
| SummEval relevance | Spearman | 0.563 | 0.534 | 0.526 |
| SummEval consistency | Spearman | 0.537 | 0.550 | 0.555 |
| SummEval (both) | score MAE vs experts | 0.94 | 0.51 | — |
| RewardBench | official score | 0.847 | 0.916 | — |
| Open-Jev (4,204) | accuracy | 66.7% | 85.0% | — |
| SCOTUS (640, ≤40k chars) | micro / macro-F1 | 66.1 / 57.3 | 64.4 / 56.4 | 72.6 / 62.1* |
| ECtHR A (300 cases) | micro / macro-F1 | 49.2 / 39.5 | 57.9 / 56.8 | — |
| ECtHR B (300 cases) | micro / macro-F1 | 72.5 / 72.2 | 71.1 / 68.0 | — |
| Mind2Web test | element accuracy | 85.6% | 92.4% | — |
The Jev numbers are third-party reports: PubMedQA and SummEval from Deußer et al., arXiv 2609.37647 (same 1,000 PubMedQA items; summary-level Spearman), SCOTUS from the chepyle/jev-test reanalysis. *That SCOTUS run used the full test split with truncation; ours skips the 760 longest opinions, so the two are not directly comparable.
Images
Shingi reads images through the Bonsai 2 27B vision projector (mmproj.gguf, included
here). Send up to 8 images per request: /v1/decisions follows SGLang's decision
endpoint, and /v1/systemone takes an optional images list.
Zero-shot:
| Suite | Items | Accuracy |
|---|---|---|
| VSR (spatial yes/no) | 300 | 82.0% |
| A-OKVQA (4-way) | 300 | 89.0% |
| VQAv2 yes/no | 300 | 89.0% |
The model was not trained on images; these come from the base model's vision with Shingi's decision training on top.
Training data
Public datasets, used under their stated licenses: Banking77, CLINC150, MMLU, HelpSteer2, HelpSteer, Measuring Hate Speech, Civil Comments, GoEmotions, LEDGAR (LexGLUE), MASSIVE, CommonsenseQA, WinoGrande, GSM8K, HelpSteer3, TabFact, TAT-QA, CaseHOLD and Unfair-ToS (LexGLUE), TimeQA, Mind2Web, MedMCQA and MedQA. Several came through the jev-bench repackaging.
About a third of the training tokens are model-generated: a private synthetic dataset of decision tasks, plus the Apache-2.0 teacher data from Mapika/decider.
Limitations
- English only.
- Much slower on a Mac than on an NVIDIA GPU: on an M4 Pro, a short decision takes about 1.2 s and a 1,100-token one about 12.6 s. MLX was measured alongside Metal and was about 10% slower on the M4, so the Mac build uses Metal.
License
Apache-2.0. Derived from Bonsai 2 27B by Prism ML (Apache-2.0), which descends from Qwen3.8-27B; see NOTICE. The training datasets keep their own licenses.