StartLux-Decision-27B

StartLux-Decision-27B answers typed questions about a state: pick one of several options, yes or no, or a rating on a scale. The state can be text, JSON or images, up to 262,144 tokens (256K). Every question comes back with a probability for each option. Requests and responses use the TypeSafe /v1/systemone format, so clients written for Jev work unchanged.

Sizes: 0.8B · 2B · 4B · 9B · 27B · 35B-A3B

Results

StartLux-Decision-27B
Decision Index 0.2.1	63.88
Decision Index 0.2	59.54
JevBench public, correct of 231	208
Intern-Decision, average accuracy over seven suites	91.82
Latency, one request with three questions	102.3 ms
Latency, one yes/no question	50.7 ms
Input	text, JSON or images, up to 262,144 tokens

Latency is end to end over HTTP on one H200 in bf16, one request at a time.

Compared with other decision models

Model	JevBench public, of 231	Intern avg	DI 0.2 / 0.2.1	Latency, 3 questions
StartLux-Decision-35B-A3B	210	92.29	57.24 / 61.55	52.5 ms
StartLux-Decision-27B	208	91.82	59.54 / 63.88	102.3 ms
StartLux-Decision-9B	201	91.08	54.37 / 58.63	35.7 ms
StartLux-Decision-4B	204	91.17	48.38 / 52.75	26.0 ms
Intern-Decision-4B	201	90.02	35.90 / 37.81	44.2 ms ¹
JevK5	200	85.16	36.44 / 38.81
Jev 1.13	199	88.74	51.67 / 57.91	64.0 ms ²
StartLux-Decision-2B	196	88.46	40.72 / 44.19	15.5 ms
SemIf	187	84.23	25.70 / 25.94
Intern-Decision-2B	180	84.68	19.49 / 19.38	33.3 ms ¹
StartLux-Decision-0.8B	179	85.03	35.57 / 38.86	12.2 ms
Intern-Decision-0.8B	163	79.38	11.32 / 11.94	34.0 ms ¹
Laya	130	57.77	5.51 / 6.04

JevBench public counts the correct answers on the 231 public items in the Intern-Decision bundle; Intern avg is the average accuracy over its seven suites; DI is the Decision Index under both editions, the public board's values for the other systems. A blank cell means the number is not published. StartLux-Decision latencies are for one H200, with the three questions answered in one forward pass. ¹ Intern-Decision's own measurement on an RTX 4090. ² The server time the TypeSafe API gateway reports for the same request, mean of 100, network left out as in ours; Intern-Decision reports 109.7 ms end to end.

Fast inference

The folder ships its own inference package, startlux_decision/, which is the fast path:

• requirements.txt installs the fast kernels, flash-linear-attention and causal-conv1d, and python -m startlux_decision.check . confirms they are active. Without them transformers falls back to a path more than ten times slower, and the server refuses to start on a GPU.
• All questions of a request run in one forward pass, and the server records CUDA graphs at start-up and replays them for short requests: on one H200 a request with three questions takes 102.3 ms end to end and a single yes/no question 50.7 ms.
• For bulk work, decide_batch batches the questions of many requests together, which is several times faster than sending them one at a time.

Images and long inputs

The weights include a vision tower, and the inference package uses it: a request can carry images as part of its evidence, images=[...] in Python (PIL images, file paths, encoded bytes, base64 strings or data URIs) or "images": [...] over HTTP (base64 strings or data URIs). <image> in a string state marks where each image goes. Prompts can run to 262,144 tokens (256K), the model's native context: a long state is read once, in chunks, and every question of the request branches off it. The MLX backend and the GGUF files read text only. confidence follows TypeSafe's definitions (for a choice, (p_max − 1/n) / (1 − 1/n)); the top probability is in probabilities.

Usage

hf download startlux-models/StartLux-Decision-27B --local-dir StartLux-Decision-27B
cd StartLux-Decision-27B
pip install -r requirements.txt
python -m startlux_decision.check .                      # must print "fast kernels: active"
python -m startlux_decision.server --model . --port 8090

curl -s localhost:8090/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": {"ticket": "I was charged twice for order #4411 and the app still shows it as unpaid."},
  "questions": {
    "team":   {"type": "choice", "instructions": "Which team should handle this ticket?",
               "criteria": {"billing": "Payments, refunds and invoices",
                            "shipping": "Delivery and tracking",
                            "technical": "App, login and account problems"}},
    "urgent": {"type": "noul", "instructions": "Should this ticket be answered today?"},
    "severity": {"type": "score", "instructions": "How severe is the impact?",
                 "criteria": ["cosmetic", "annoying", "blocks the customer"]}
  }
}'

Or in Python, from the same folder:

from startlux_decision import StartLuxDecision

m = StartLuxDecision(".")
answers, usage = m.decide(state, questions)
many = m.decide_batch([(state, questions), ...])

With images, from the same folder:

answers, usage = m.decide("Photo taken at delivery: <image>",
                          {"damaged": {"type": "noul", "instructions": "Is the parcel damaged?"}},
                          images=["parcel.jpg"])

License

The model weights are released under CC BY-NC 4.0: free for research and other non-commercial use, with attribution. Commercial use requires a separate license from StartLux Labs; contact contact@startlux.com. The inference code in startlux_decision/ is Apache-2.0. See LICENSE and NOTICE.