Julia 1.5

A 144-million-parameter model for text decisions, with stronger rule following and less sensitivity to option order than Julia 1.

Download Julia 1.5 on Hugging Face

Two engraved circles flank a blue diamond on ivory paper

How it works

Pass a state in text or JSON, a question, and the allowed answers. The model compares the options in one forward pass and returns a probability for each. An application can use the result to route a support ticket, check whether a request meets a policy, or score it against a rubric.

Question typeWhat you supplyWhat you get
choice2–20 named options with descriptionsThe selected key and each option’s probability
noulA yes/no question, optionally with descriptions for true and falseProbability of “yes”
score2–20 ordered rubric levelsExpected score, most likely level, and probabilities

Julia 1.5 uses a 22-layer mmBERT-small encoder and a decision head that compares the options. The 1,024-token budget is shared across the state, question, and options. Each option can take up to 48 tokens, and the question and options together have a 512-token limit. The runtime flags truncated states. Use strict mode when the whole state must fit.

Results

Julia 1.5 improves across all three core SLCMB question types in English and Portuguese. SLCMB points are a composite score. The other figures are rates on the named tests. Evaluation data · SLCMB protocol

EvaluationJulia 1Julia 1.5
SLCMB 3.0, full 920-question set33.346.2
Written rules followed29.6%87.4%
Negated rules followed27.6%94.3%
Same decision after options are reordered59%87%
Public classification tests in five languages67.0%73.0%

Some tasks regress. On the SLCMB Emotion track, Julia 1 scored 91.6 and Julia 1.5 scored 66.0. Exam-style world knowledge remains limited. Compute arithmetic and tight numerical comparisons in application code and send the result as part of the state. Validate probabilities on your own data before using thresholds to trigger actions.

On an NVIDIA L40S, the model card reports about 15 ms for one request and about 3,700 rows/s in batches with the FP16 encoder. Its 8-vCPU FP32 result was about 250 ms for one 325-token request and about 5 rows/s. These are hardware- and workload-specific measurements. Model card and runtime benchmarks

Run locally

Use Python 3.10+ and PyTorch on CPU or CUDA. The Hugging Face repository includes the FP32 weights (577 MB), tokenizer, runtime, metrics and provenance under Apache 2.0.

python -m pip install huggingface_hub
hf download SupersonicLabs/Julia-1.5 --local-dir Julia-1.5
python -m pip install -e ./Julia-1.5/runtime
from julia15 import Runtime

model = Runtime("Julia-1.5")
result = model.predict_questions(
    state={"ticket": "I was charged twice for the same order."},
    questions={
        "team": {
            "type": "choice",
            "instructions": "Which team should handle this ticket?",
            "criteria": {
                "billing": "Duplicate charges, payments and refunds",
                "technical": "Login issues and app errors",
            },
        },
    },
)
print(result["answers"]["team"])

Start the local HTTP runtime with python -m julia15 serve --port 8080. Use POST /v1/predict for one request, /v1/predict_batch for a batch, or /v1/logits for raw option scores. For choices where option order matters, try views=3 to evaluate several orderings. Runtime documentation

Flash and Nano

Julia 1.5 Flash

A blue diagonal stripe crosses a pale engraved parallelogram

Julia 1.5 Flash is a private preview. It keeps the question types, 1,024-token context and runtime of Julia 1.5, with 296 MB of weights versus 577 MB for the standard release.

Julia 1.5 Nano

Julia 1.5 Nano has no public release. For an ESP32-S3 text runtime, see Julia 2 Nano.

Migrating from Julia 1

We’re retiring Julia 1 on the hosted API. Requests to the Julia 1 endpoint now route to Julia 1.5 with the same request format. Local installations switch by downloading the new weights and runtime. Recheck your own decision thresholds, especially if your workload uses the Emotion track or contexts longer than 1,024 tokens. Julia 1 research archive