sergolsub/mlx-startlux-v1.1-9B

StartLux-Decision v1.1 9B (MLX 8-bit mirror)

Byte-identical mirror of the official MLX 8-bit release startlux-models/StartLux-Decision-v1.1-9B-Instant-MLX-8bit for Apple Silicon (mlx_lm); the upstream authors quantized it.

Origin and license

• Original weights: https://huggingface.co/startlux-models/StartLux-Decision-v1.1-9B-Instant-MLX-8bit
• License: cc-by-nc-4.0 — carried over from the upstream weights; if anything conflicts, the upstream repository's LICENSE prevails.
• Conversion: 8-bit quantization for MLX only; no retraining, no architecture changes.

Non-commercial. CC BY-NC 4.0: commercial use is not permitted; keep the upstream attribution when redistributing.

Measured (Mac mini M4 Pro 48 GB, MLX 8-bit)

• Standard stand — load 4.98 s · peak RSS 11.21 GB
• Text 512 tokens (448 tok) — avg of 5 after 3 warm-up: binary 1210 ms · multichoice 1299 ms · score 1303 ms · peak RSS 11.21 GB
• Box: not supported
• On disk 10.44 GB · load 8.7 s · peak resident memory 4.91 GB — weights stream from disk (RSS is the working set, not the model footprint)
• Decision latency: median 521 ms · sustained 3.58 short decisions/s
• Quality after 8-bit quantization (our eval): accuracy 87.8% (8-bit) vs 87.8% (bf16) · noul 95.0% vs 95.0% · choice 80.0% vs 80.0% · score 84.0% vs 84.0%
• Calibration PPL 1.745 (8-bit) vs 1.745 (bf16)
• Decision Index 0.4.1: 60.95

Usage

# official Instant MLX release — stock mlx_lm loads it as a plain backbone:
from mlx_lm import load, generate
model, tokenizer = load("sergolsub/mlx-startlux-v1.1-9B")
print(generate(model, tokenizer, prompt="...", max_tokens=256))
# decisions (System One): the official StartLux-Decision package, or our
# served decisions over HTTP (start the server first: ./run_local_all.sh):
# one task type per call:
import json, urllib.request
def ask(state, questions):
    req = urllib.request.Request(
        "http://127.0.0.1:8090/v1/systemone",
        data=json.dumps({
            "model": "startlux-v1.1-27B", "state": state, "questions": questions
        }).encode(), headers={"Content-Type": "application/json"})
    return json.loads(urllib.request.urlopen(req).read())["answers"]

print(ask("I was charged twice for order #4411 and the app still shows it as unpaid.",
          {"refund": {"type": "noul", "instructions": "Does the customer want a refund?"}}))
print(ask("Order #7821 shipped on Monday but tracking has not updated for 5 days.",
          { "team": {"type": "choice", "instructions": "Which team should handle this?", "criteria": {"billing": "Payments and refunds", "shipping": "Delivery and tracking"}} }))
print(ask("Export failed: disk quota exceeded on the worker.",
          { "severity": {"type": "score", "instructions": "How severe is the impact?", "criteria": ["cosmetic", "annoying", "blocks"]} }))

Files

• *.safetensors — 8-bit weights (MLX layout)
• config.json, tokenizer*, chat_template.jinja — standard transformers/MLX set
• *_mlx.py / decision_config.json / joint_config.json — decision/joint-head inference glue (present when the model has a decision mode)

Checksums (SHA-256)

SHA-256 of every file in this repository except this README itself: LFS digests pulled from the HF metadata, small files (config, tokenizer, …) hashed on demand — the safetensors shards included. Re-check any download with sha256sum <file> and compare.

File	SHA-256	Size
.gitattributes	9c312bebcc1fe2dcbc46d5a6bed58b112fbe55b906b339c2cf7320afb3b3e0f9	1,632 B
LICENSE	41003d4a74749c0220e33dd415042164b5a1093ed401f36277234f772d22d3d0	19,347 B
NOTICE	ec7d09e98f04667acf3a33358ee4e830f6ee3b41251af388035aec40d4e47fe2	12,330 B
chat_template.jinja	a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715	7,756 B
config.json	b00915f594eee7012231e8dcde19602aa51a88e38120a5d4e98bc965656e18e5	3,376 B
decision_config.json	7941d0886698e7f9ed5591777422fb718c33f5c730e5c14c64b653ec534ab2ca	521 B
figures/score_vs_size.png	a2ab77cac1922a33414c88bc68e030310889e679c4d9dfde8bb3dd4a067042a9	347,430 B
model-00001-of-00002.safetensors	625e24ca13cf6c9ce98ca3c63f531c29962523365b3fbbd5ae4d644f585acfa7	5.34 GB
model-00002-of-00002.safetensors	22b389a4abc6f830ed15acc4a94c53d9e851cd44c8ce2bed313756c6ab2115ab	5.09 GB
model.safetensors.index.json	87d95b037f57d101448b262059b7d28d65d55f60231100cdb1827d7cd44202cd	123,593 B
preprocessor_config.json	27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516	390 B
tokenizer.json	5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42	13 MB
tokenizer_config.json	316230d6a809701f4db5ea8f8fc862bc3a6f3229c937c174e674ff3ca0a64ac8	16,710 B
video_preprocessor_config.json	7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13	385 B