infosave/d1-2B-cmf

d1-2B — CMF

Liquid AI d1-2B — a 2B decision model (System One) on LFM3-VL-2B: named questions over a state (text, JSON or pictures) answered from one forward pass as typed answers with calibrated probabilities, with no generated tokens — packaged as single CMF files. They run with cortiq, a Rust inference engine with no Python and no ML framework: NVIDIA, AMD and Intel GPUs through Vulkan/DX12, Apple silicon through Metal, and the CPU.

English · Русский · 中文

Quick start

cargo install cortiq-cli
hf download infosave/d1-2B-cmf d1-2b-q8_2f.cmf --local-dir .
cortiq system-one d1-2b-q8_2f.cmf \
  --state "I was charged twice and my parcel never arrived. Please refund one charge." \
  --questions '{"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
                "team": {"type": "choice", "instructions": "Which team should handle this?",
                         "criteria": {"billing": "Charges, refunds, invoices",
                                      "shipping": "Delivery and parcels",
                                      "technical": "App or site faults"}}}'

{"answers": {"refund": {"type": "noul", "noul": 0.9972},
             "team": {"type": "choice", "choice": "billing", "confidence": 0.7028,
                      "probabilities": {"billing": 0.7028, "shipping": 0.2477, "technical": 0.0495}}},
 "usage": {"input_tokens": 88, "output_tokens": 0}}

Use cortiq 0.8.50 or later; prebuilt binaries for Linux, macOS and Windows are on the releases page. --state and --questions take text or @file (a .json state file is read as JSON); --image adds a picture (path, data URL or HTTP(S) URL, repeatable).

Files

file	weights	vision encoder	size	answers = original	recommended for
d1-2b-q8_2f.cmf	8-bit	16-bit	2,135,818,786 B (2.14 GB)	19 of 19	default
d1-2b-q4tp.cmf	4-bit	16-bit	1,306,104,354 B (1.31 GB)	19 of 19 (one estimate 2 steps off)	the smallest memory footprint

Answers = original: the choice, level, span or value of every question of four test requests (a support ticket, an invoice, a JSON record and a picture with counts and boxes) against the bf16 checkpoint run in float32 by Transformers; the largest probability difference is 0.014 (q8_2f) and 0.07 (q4tp). An f16 conversion matches Transformers float32 to about 1e-6 on every probability. The calibration the release ships (calibration.json) is in the file and applied to every answer.

Performance

One request at a time through cortiq serve, after the first call:

hardware	file	1 question (43 tokens)	3 questions (123 tokens)	5 questions over a ~200-token state
NVIDIA A40 (Vulkan)	q4tp	45 ms	64 ms	100–115 ms
NVIDIA A40 (Vulkan)	q8_2f	56–75 ms	70 ms	102–128 ms
Apple M4 (Metal)	q4tp	85 ms	160 ms	255–310 ms

The whole forward pass — every layer over the packed state and questions, then the vocabulary head over the answer slots — runs in one GPU submission. A picture adds its vision encoder (0.5 s for 640×480 on the M4).

Questions

type	question	answer
noul	yes or no	noul: P(yes)
choice	one of the options criteria names ({name: description})	choice, confidence, probabilities
choice with "multiple": true	any of the options	choices, probabilities (each option's own P(yes))
score	one of 2 to 10 ordered levels (criteria lowest first)	score (the median level), confidence, probabilities, legend
text, number, date	a value read from the state	the value, span, found
number with range (and step)	an estimate	number (the median), interval (p10, p90)
box, boxes (max)	one object, or every instance, in the picture	box (top, bottom, left, right as fractions), found, confidence, or boxes and count

A value or a box whose found is under 0.5 is null. Up to 16 questions are read together after one state (the release's set size); more run in further passes over the same state.

cortiq system-one d1-2b-q8_2f.cmf --image cats.jpg \
  --questions '{"cats": {"type": "number", "instructions": "How many cats are there?", "range": [0, 10], "step": 1},
                "remote": {"type": "box", "instructions": "Where is the remote control on the left?"}}'

Server

cortiq serve d1-2b-q8_2f.cmf --port 8080

curl http://localhost:8080/v1/system_one \
  -H "Content-Type: application/json" \
  -d '{"state": "ACME SA\nInvoice 2026-117\nTotal due: 1 234,50 EUR",
       "questions": {"total": {"type": "number", "instructions": "How much is due?"}},
       "media": []}'

media takes pictures (a path, an HTTP(S) or data: URL, or {"type": "image_url", "image_url": {"url": ...}}). Video and batched states (system_one_batch) are not supported yet.

Hardware

NVIDIA / AMD / Intel (Vulkan, DX12). Both files fit any card with 2.5 GB of free memory; on headless Linux install libvulkan1 libglvnd0 libegl1 libgl1 libglx0 and set XDG_RUNTIME_DIR=/tmp. Apple silicon: the Metal backend is the default. CPU only: CMF_GPU=0.

Verify the download

sha256sum -c d1-2b-q8_2f.cmf.sha256
cortiq verify d1-2b-q8_2f.cmf

file	sha256
d1-2b-q8_2f.cmf	afee01128b798277a49ddd97ebfa37bb8f94cca16e245f4f40065cab1a3c2272
d1-2b-q4tp.cmf	65ce90e5df70f591bc1594c40c9425a39770d257433a51ee30b01892eb638e04

Reproduce

cortiq convert --model LiquidAI/d1-2B --quant q8_2f \
  --output d1-2b-q8_2f.cmf --also-output q4tp=d1-2b-q4tp.cmf

Scripts (Transformers reference, answer parity, latency): scripts/lfm3.

License

These files are a modified form of Liquid AI's d1-2B: the weights were converted to CMF and the language model quantized (8-bit q8_2f, 4-bit q4tp) by cortiq. They remain under the LFM Open License v1.0, a copy of which is in this repository (LICENSE); among its terms, commercial use by an entity with annual revenue of US$10 million or more is not licensed. See the upstream card.

---

Документация на русском

d1-2B от Liquid AI — модель решений (System One) на 2 млрд параметров поверх LFM3-VL-2B: именованные вопросы к состоянию (текст, JSON или изображения) получают типизированные ответы с откалиброванными вероятностями за один прямой проход, без генерации токенов, — в виде файлов CMF для движка cortiq на Rust без Python.

cargo install cortiq-cli
hf download infosave/d1-2B-cmf d1-2b-q8_2f.cmf --local-dir .
cortiq system-one d1-2b-q8_2f.cmf --state "Мне дважды списали деньги, посылка не пришла." \
  --questions '{"refund": {"type": "noul", "instructions": "Клиент просит возврат?"}}'

файл	веса	размер	ответы = оригинал	назначение
d1-2b-q8_2f.cmf	8 бит	2.14 ГБ	19 из 19	по умолчанию
d1-2b-q4tp.cmf	4 бита	1.31 ГБ	19 из 19 (одна оценка на 2 шага)	минимум памяти

Нужен cortiq 0.8.50 или новее. Типы вопросов: noul (да/нет), choice (один или несколько вариантов), score (уровни), text/number/date (значения из состояния), number с range (оценка с интервалом), box/boxes (рамки на изображении). Весь проход — слои над упакованным состоянием и вопросами и голова словаря над слотами ответов — идёт одной отправкой на GPU: на NVIDIA A40 один вопрос отвечается за 45 мс (q4tp), на Apple M4 (Metal) — за 85 мс. cortiq serve отдаёт POST /v1/system_one (state, questions, media). Видео и пакетные запросы пока не поддерживаются. Файлы — изменённая форма d1-2B от Liquid AI по LFM Open License v1.0 (копия — LICENSE); коммерческое использование организацией с годовой выручкой от 10 млн долларов США лицензией не покрывается.

中文说明

d1-2B 是 Liquid AI 基于 LFM3-VL-2B 的 2B 决策模型（System One）：对一个状态（文本、JSON 或图片）提出若干具名问题，一次 前向即可得到带校准概率的类型化答案，不生成任何 token。本仓库提供可由 Rust 推理 引擎 cortiq（0.8.50 及以上）加载的 CMF 文件。

cargo install cortiq-cli
hf download infosave/d1-2B-cmf d1-2b-q8_2f.cmf --local-dir .
cortiq system-one d1-2b-q8_2f.cmf --state "我被重复扣款，包裹也没到。" \
  --questions '{"refund": {"type": "noul", "instructions": "客户是否要求退款？"}}'

文件	权重	大小	答案与原模型一致	适用
d1-2b-q8_2f.cmf	8 位	2.14 GB	19/19	默认
d1-2b-q4tp.cmf	4 位	1.31 GB	19/19（一个估计值差 2 档）	内存最小

整个前向（打包后的状态与问题经过所有层，再在答案位置上计算词表头）在一次 GPU 提交中完成：NVIDIA A40 上单个问题 45 ms（q4tp），Apple M4（Metal）上 85 ms。 cortiq serve 提供 POST /v1/system_one。暂不支持视频与批量请求。这些文件是 Liquid AI d1-2B 的修改形式，仍受 LFM Open License v1.0 约束（副本见 LICENSE）；年收入 1000 万美元及以上的实体进行商业使用不在许可范围内。