Request access to Myles-G 12B
Myles-G 12B is available to approved testers for internal evaluation under the Myles Evaluation Licence. Tell us who you are and what you want to evaluate; every request is reviewed and approved by hand.
Log in or Sign Up to review the conditions and access this model content.
Myles-G 12B
Myles-G 12B is a decision model: give it a situation and typed questions (yes or no, one choice of several, a level on a scale) and it reads every option in one forward pass and returns a calibrated probability for each, with no generated text. It runs on one workstation GPU you own, or through the Launchables decision API. Built by Launchables in the UK.
Status: evaluation build for approved testers; not for production; provenance review in progress. These are the exact weights behind the Myles-G 12B decision API since 12 October 2026, with the same readout settings (
runtime.json). The figures below are our own measurements; this build may be replaced by a later one. See "Training data" for what this build was trained on.
At a glance
| JevBench public decision set | 201 of 231 (87.0%) through the served API with the benchmark's harness; 203 of 231 offline (bf16) |
| Output | A probability for every option; a decision and a confidence you can set a threshold on |
| Time per decision | about 50 ms for a short decision on one NVIDIA RTX A6000 (bf16, vLLM) |
| Runs | Entirely on your own hardware; no data leaves your premises |
| Weights | bf16 safetensors, 23 GB, merged (no adapter to apply) |
| Build | 12 October 2026 |
| Access | Gated; requests are approved manually |
Our own measurements on the 231 public JevBench items, 12 October 2026, using the benchmark's published scoring code: 201 through the served endpoint (easy 48 of 48, original 70 of 72, hard 83 of 111), 203 offline on one NVIDIA H100 in bf16. Not an official placement; an official evaluation was requested on 12 October 2026. JevBench's official ranking also weighs sealed items, calibration, speed and cost.
Access
This repository is gated. Request access above; every request is reviewed and approved by hand. Approved access covers downloading and running Myles-G 12B on hardware you control for internal evaluation (see LICENSE). Production or commercial use needs a separate agreement: email info[at]launchables[dot]ai.
What is in this repository
| File | What it is |
|---|---|
model-0000*-of-00005.safetensors, config.json, tokenizer and chat template |
the Myles-G 12B weights (Gemma 4 12B with Launchables' adaptation merged in), bf16 |
runtime.json |
the readout settings the API serves with: layout cygnet, one temperature (1.0533), 32,768-token context, states over 20,000 characters cut to their beginning; it pins the weights to an exact commit |
runtime/ |
the Myles decision runtime (Python): starts vLLM on these weights and serves the decision API on top |
LICENSE, THIRD_PARTY_NOTICES.md, LICENSE-GEMMA-APACHE-2.0.txt |
the Myles Evaluation Licence and the third-party terms |
Hardware
- As served: one 48 GB GPU (for example NVIDIA RTX A6000 or L40S) in bf16.
- 24 GB GPU (for example RTX 4090): add
--quantization fp8. The temperature was fitted in bf16, so fp8 is a slightly different model; on our RTX 4090 the fp8 build read 202 of 231 public items, and the commands below were tested that way. - Linux with a recent NVIDIA driver (CUDA 13), Python 3.12, about 25 GB of disk for the weights plus about 10 GB for the Python environment.
Run it
You need a Hugging Face access token from the account that was granted access.
export HF_TOKEN=hf_... # read token of the approved account
pip install -U huggingface_hub # provides the `hf` command
hf download launchables/myles-g-12b --include "runtime/*" runtime.json --local-dir myles-g-12b
cd myles-g-12b
uv venv -p 3.12 .venv # or python3.12 -m venv .venv
uv pip install -p .venv/bin/python "./runtime[serve]"
export MYLES_RUNTIME_API_KEY=choose-a-secret # requests must carry it as a bearer token
export VLLM_USE_FLASHINFER_SAMPLER=0 # not needed when the CUDA toolkit (nvcc) is installed
.venv/bin/myles-runtime --config runtime.json --host 127.0.0.1 --port 8010
# on a 24 GB card add: --quantization fp8
myles-runtime downloads the weights pinned in runtime.json (about 23 GB) into ./models, starts vLLM on a local port and serves the decision API on port 8010. The first start takes a few minutes; it is ready when curl http://127.0.0.1:8010/health answers.
curl -s http://127.0.0.1:8010/v1/systemone -H "Authorization: Bearer $MYLES_RUNTIME_API_KEY" -H "Content-Type: application/json" -d '{
"state": "The customer says they were charged twice for order 4471 and wants the extra payment back.",
"questions": {
"intent": {"type": "choice", "instructions": "What does the customer want?", "criteria": {"refund": "Money back", "track_order": "Where is my order"}},
"duplicate": {"type": "noul", "instructions": "Is the customer reporting a duplicate charge?"},
"urgency": {"type": "score", "instructions": "How urgent is this?", "criteria": ["Not urgent", "Soon", "Today"]}
}}'
Each question comes back with a probability for every option: noul returns the probability of yes, choice the chosen label and every option's probability, score the expected level, the most likely level and every level's probability. usage reports input tokens; nothing is generated. Up to 16 questions per request; POST /v1/decide is the same handler.
How it decides
Each question becomes one prompt: the situation, the question and the options as a lettered list. The answer is read in one pass at the first generated position, restricted to the option letters, renormalised over the options shown, and divided by a temperature fitted on Launchables' own held-out rows (never on benchmark items). One forward pass per question; no generated text.
Training data
Launchables adapted Gemma 4 12B with low-rank training on public, human-labelled decision datasets and code-generated decisions with computed answers, each with its licence recorded (see THIRD_PARTY_NOTICES.md). These weights are the average of two such adapted models, merged into the base. Labels are each dataset's own or computed by code; no model labelled training rows. JevBench items are never used for training or for fitting the readout, and the training data is checked for overlap with them (no overlap with the 231 public items).
Provenance, stated plainly. This build's training data includes the following, which are under review before any wider or commercial release:
- About 7,800 code-generated rows whose phrasing templates were written with the help of a hosted AI assistant (Launchables' skills data). The labels are computed by code; the templates are being replaced with human-written ones.
- Rows from a public decision-model dataset released under CC0 1.0 whose generator lineage we have not yet confirmed.
- FaithDial, whose evidence text is Wikipedia under CC BY-SA (share-alike).
- QASC, whose wrong answer options were selected with the help of a model (the questions and answers are human-written).
Limitations
- Benchmark scores are on public items; results on sealed items and on your own decisions may be lower. Measure on your own held-out cases before relying on it.
- Confidence is calibrated on held-out data, not guaranteed; set thresholds per decision and keep a person in the loop for the rest.
- English only in this build.
- At most 26 options per question (one letter each).
Lineage and licence
Myles-G 12B is adapted by Launchables from Google's Gemma 4 12B (google/gemma-4-12B-it, Apache License 2.0). The Myles Evaluation Licence (see LICENSE) covers the release: internal evaluation by approved testers. Gemma 4 remains under the Apache License 2.0 (see THIRD_PARTY_NOTICES.md and LICENSE-GEMMA-APACHE-2.0.txt). Gemma is a trademark of Google LLC.
More
Benchmarks: launchables.ai/benchmarks. Contact: launchables.ai/contact.
- Downloads last month
- 4