Results
This is how linden-1 compares against TypeSafe Jev 1.14 and GPT-6 Luna on a wide range of decision model benchmarks.
Table with columns: Benchmark, linden-1, TypeSafe Jev 1.14, GPT-6 Luna| Benchmark | linden-1 | TypeSafe Jev 1.14 | GPT-6 Luna |
|---|
| Decision Index 0.3, public tests | 71.3 | 60.5 | 47.3 |
| JevBench (public split, 231) | 92.2 | 88.3 | 89.6 |
| Amazon ESCI (5,000, macro-F1) | 63.6 | 56.9 | 53.7 |
| New Yorker caption contest (528) | 87.1 | 72.0 | 69.5 |
| PubMedQA (1,000) | 81.3 | 79.4 | 71.9 |
| Global-MMLU-Lite, non-English (8,800) | 84.2 | 84.7 | 80.4 |
| When2Call (3,652) | 91.2 | 81.9 | 69.1 |
| Top-1 calibration error (ECE), mean of the six, lower is better | 0.030 | 0.067 | 0.093 |
We ran all three models ourselves on the same items with the same scorer: Jev 1.14 through TypeSafe's API and GPT-6 Luna through OpenAI's Decisions API. A question a model refuses or cannot answer counts as wrong. Each question is asked in each provider's native format (yes/no, choice or score); prompts, options and states are identical. The Decision Index row is the full public suite of the Decision Index 0.3 edition, scored with the kit's own scorer; its benchmarks score skill above a baseline, so a model that leans to one answer scores near zero on them.
Latency
Time to answer one request, sent one at a time: median and 90th percentile of 15 requests per case. Every request
starts each text field of its state with a new random number, so no prefix cache can answer it.
Table with columns: Request, linden-1, TypeSafe Jev 1.14, GPT-6 Luna| Request | linden-1 | TypeSafe Jev 1.14 | GPT-6 Luna |
|---|
| Moderation: a forum comment (~150 tokens), 1 yes/no question | 53 ms (p90 54) | 165 ms (p90 187) | 334 ms (p90 415) |
| Tool routing: a short chat and 5 tools (~400 tokens), 5 yes/no questions | 103 ms (p90 103) | 158 ms (p90 177) | 333 ms (p90 513) |
| Support triage: a customer email (~1,000 tokens), choice, score and yes/no | 159 ms (p90 160) | 163 ms (p90 188) | 284 ms (p90 463) |
| Contract review: a services agreement (~3,400 tokens), choice and yes/no | 248 ms (p90 249) | 172 ms (p90 181) | 344 ms (p90 397) |
| Incident report: a postmortem (~8,500 tokens), 1 yes/no question | 367 ms (p90 368) | 219 ms (p90 228) | 275 ms (p90 341) |
linden-1 was measured on one NVIDIA B300 serving it alone, with the client on the same host. Jev 1.14 and GPT-6 Luna
were measured through their public APIs from a client about 15 ms from their endpoints, so their times include a short
network hop and their API layers. linden-1 is fastest on short requests and steady at every size; Jev 1.14 is faster on
inputs from about 1,000 tokens, where linden-1 reads short inputs twice for accuracy and its cost grows with input
length.
Using linden-1
linden-1 is available on the Gradiated API as gradiated/linden-1:
curl https://api.gradiated.com/v1/systemone \ -H "Authorization: Bearer $GRADIATED_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "gradiated/linden-1", "state": {"ticket": "I was charged twice. Please fix this today."}, "questions": { "category": {"type": "choice", "instructions": "Choose the department for this ticket.", "criteria": {"billing": "Charges or payments", "technical": "Something is broken", "other": "Anything else"}}}}'
Try it out and let us know how you find it.
Training
- Initial checkpoint: KnowLine-31B-Gen1.5 (PelaAI, Apache-2.0), a decision fine-tune of Gemma 4 31B.
- Stage 1: 952k rows, 1,155 steps (0.25 epoch): procedural typed decisions, the commercial-licence rows of
tasksource's Jev-format decisions, the official train splits of 14 Decision Index benchmarks and sn92's sets.
- Stage 2: 563k rows, 1,000 steps (0.35 epoch) from stage 1, aimed at what it performed poorly at: temporal and numeric reasoning
(TempReason, TimeQA, TAT-QA, FinQA), long policies and contracts (ShARC, CUAD, MAUD, LexGLUE), preference judging
(HelpSteer2/3, OASST2, Stack Exchange), dialogue state (MultiWOZ). About a sixth of the rows were replayed from stage 1 so it kept what it had learned.
- Setup: full fine-tune, BF16 compute with FP32 master weights, soft cross-entropy on the option distribution plus 0.5 × Brier score. About 101 GPU-hours across B200/B300s.
- Precision: trained in BF16 and quantized to FP8 (33 GB) for release; the BF16 weights are on the
bf16 branch.
- Blend: the released weights are 0.5 × the initial checkpoint + 0.5 × the stage 2 checkpoint.
Data and licences
All training data is public. We excluded sources whose licence forbids commercial
use or is unstated, and removed rows written by OpenAI, Anthropic or Google models from mixed-author sources. Every source was decontaminated (exact match and word 13-gram overlap) against the Decision Index 0.3 test rows, the kit's
test pools, and the JevBench and Gemini launch panel items.
Sources: procedural typed decisions; sn92 (generalist and specialists); tasksource Jev-format decisions (commercial
rows only); the train splits of API-Bank, BANKING77, CLINC150, SGD, ARC, WinoGrande, HellaSwag, GSM8K, ChessBench,
Amazon ESCI, HoVer, When2Call and the New Yorker caption contest; TAT-QA, FinQA, ConvFinQA, DROP, TimeQA, TempReason,
ShARC, CUAD, MAUD, LexGLUE, CondaQA, HelpSteer2, HelpSteer3, OASST2, Stack Exchange preferences, MASSIVE, sarcasm
headlines, IBM claim stance and MultiWOZ 2.2.
Their licences are permissive (Apache-2.0, MIT, BSD, CC-BY) or share-alike (CC-BY-SA, for example DROP, Stack Exchange
preferences, TempReason and ShARC). We release the weights under Apache-2.0, the licence of the base model, and credit
every source here.
Amazon ESCI, When2Call and New Yorker are among the benchmarks whose train splits we used; their test rows were
decontaminated, but these are in-distribution results.
Limitations
- Yes/no lean. With no evidence either way,
noul answers lean towards yes.
- Languages. Trained and evaluated mostly in English, so performance outside of that has not been measured.
Where did the name come from?
linden-1 is named after the linden tree. In medieval Germany, village courts and
assemblies met outdoors under a linden in the middle of the village. People brought their disputes, the case was heard, and a judgment was given there and then.
Records of the time say judgments were given sub tilia, "under the
linden", and some of these court lindens (Gerichtslinden) still stand.
Licence
Apache-2.0. linden-1 contains weights from KnowLine-31B-Gen1.5 (Apache-2.0, PelaAI), which derives from Gemma 4
(Apache-2.0, Google).