infosave/LFM3-1.7B-Base-cmf

LFM3-1.7B-Base — CMF

Liquid AI LFM3-1.7B-Base — the pre-trained checkpoint behind LFM3-1.7B (Liquid's tri-hybrid architecture: 24 layers, 13 gated short-convolution, 6 full-attention and 5 sliding-window attention; 262,144-token context), meant as a starting point for fine-tuning and research — 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, and the CPU. For chat, tool calling and agents Liquid recommends the instruct model, LFM3-1.7B.

English · Русский · 中文

Quick start

cargo install cortiq-cli
hf download infosave/LFM3-1.7B-Base-cmf lfm3-1.7b-base-q8_2f.cmf --local-dir .
cortiq ppl lfm3-1.7b-base-q8_2f.cmf -f text.txt --tokens 4096
cortiq run lfm3-1.7b-base-q8_2f.cmf --prompt "What is the capital of France?" --max-tokens 512

Use cortiq 0.8.49 or later; prebuilt binaries for Linux, macOS and Windows are on the releases page. The base model saw chat data with and without thinking traces in mid-training, in the same template as LFM3-1.7B, so cortiq run and cortiq serve work with it; --no-think asks for a direct answer.

Files

file	weights	size	wiki ppl	top-1 vs original	recommended for
lfm3-1.7b-base-q8_2f.cmf	8-bit	1,756,991,387 B (1.76 GB)	9.556	98.6 %	default: near-lossless
lfm3-1.7b-base-q4tp.cmf	4-bit	927,275,931 B (0.93 GB)	9.984	87.4 %	the smallest memory footprint
bf16 original (Transformers)	—	3.46 GB	9.556	—	—

Perplexity over the first 4096 tokens of wikitext-2 test (with BOS); top-1 vs original: positions where the file's top prediction matches the bf16 checkpoint run in float32 by Transformers over the same tokens. Norms, the short-convolution taps and the per-head attention gates stay 16-bit in both files. The architecture path is the one checked exactly against Transformers on LFM3-1.7B (an f16 conversion matched float32 position for position).

Performance

cortiq bench --core --ignore-eos, one stream, cortiq 0.8.49, NVIDIA A40 (48 GB, Vulkan):

file	decode, tok/s	decode at 1k / 4k	1k / 4k-token prompt, tok/s	first token, 4k prompt
q8_2f	166	133 / 131	663 / 657	7.0 s
q4tp	228	160 / 166	624 / 622	8.2 s

Resident weights: 1.73 GB (q8_2f), 0.90 GB (q4tp).

Apple M4 (Metal), cortiq 0.8.50: q4tp decodes 84 tok/s (82 at a depth of 1k) and reads a 1k-token prompt at 565 tok/s; every layer of a token runs in one Metal command buffer.

Hardware and context

NVIDIA / AMD / Intel (Vulkan, DX12). Both files fit any card with 2 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.

The model supports 262,144 tokens; the KV cache holds 32,768 by default and CMF_MAX_SEQ=262144 raises it. KV grows by about 25 KB per token (the short-convolution layers keep a fixed state, the sliding-window layers their last 256 positions).

Verify the download

sha256sum -c lfm3-1.7b-base-q8_2f.cmf.sha256
cortiq verify lfm3-1.7b-base-q8_2f.cmf

file	sha256
lfm3-1.7b-base-q8_2f.cmf	612bc1f196da71ca3d0b65fc0cc5472500ebb05d5a5c9ab4cedc978e6f053647
lfm3-1.7b-base-q4tp.cmf	b2d2ad4217c58e2c283a4e5fc492f74bfb03e90239cd98cf61f1f1bbbd8910c5

Reproduce

cortiq convert --model LiquidAI/LFM3-1.7B-Base --quant q8_2f \
  --output lfm3-1.7b-base-q8_2f.cmf --also-output q4tp=lfm3-1.7b-base-q4tp.cmf

Scripts: scripts/lfm3.

License

These files are a modified form of Liquid AI's LFM3-1.7B-Base: the weights were converted to CMF and 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

LFM3-1.7B-Base от Liquid AI — предобученный чекпойнт, из которого сделана LFM3-1.7B (трёхгибридная архитектура Liquid: 24 слоя — 13 короткой свёртки, 6 полного внимания, 5 со скользящим окном; контекст 262 144 токена), — основа для дообучения и исследований, в виде файлов CMF для движка cortiq на Rust. Для чата, инструментов и агентов Liquid советует инструктивную LFM3-1.7B.

cargo install cortiq-cli
hf download infosave/LFM3-1.7B-Base-cmf lfm3-1.7b-base-q8_2f.cmf --local-dir .
cortiq run lfm3-1.7b-base-q8_2f.cmf --prompt "Какая столица Франции?" --max-tokens 512

файл	веса	размер	ppl wiki	top-1 к оригиналу	назначение
lfm3-1.7b-base-q8_2f.cmf	8 бит	1.76 ГБ	9.556	98.6 %	по умолчанию: почти без потерь
lfm3-1.7b-base-q4tp.cmf	4 бита	0.93 ГБ	9.984	87.4 %	минимум памяти
bf16-оригинал (Transformers)	—	3.46 ГБ	9.556	—	—

Нужен cortiq 0.8.49 или новее. На NVIDIA A40 (Vulkan) q8_2f генерирует 166 ток/с и читает промпт 4k со скоростью 657 ток/с, q4tp — 228 и 622 ток/с. Контекст до 262 144 токенов: CMF_MAX_SEQ=262144. На Apple M4 (Metal, cortiq 0.8.50) q4tp генерирует 84 ток/с. Файлы — изменённая форма LFM3-1.7B-Base от Liquid AI (конвертация в CMF и квантование cortiq) 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

LFM3-1.7B-Base 是 Liquid AI 的预训练检查点，LFM3-1.7B 即由其 训练而来（三混合架构：24 层，13 层短卷积、6 层全注意力、5 层滑动窗口注意力； 262,144 token 上下文），适合作为微调与研究的起点。本仓库提供可由 Rust 推理引擎 cortiq（0.8.49 及以上）加载的 CMF 文件。对话、工具调用与智能体请使用 LFM3-1.7B。

cargo install cortiq-cli
hf download infosave/LFM3-1.7B-Base-cmf lfm3-1.7b-base-q8_2f.cmf --local-dir .
cortiq run lfm3-1.7b-base-q8_2f.cmf --prompt "你好" --max-tokens 512

文件	权重	大小	wiki ppl	与原模型 top-1 一致率	适用
lfm3-1.7b-base-q8_2f.cmf	8 位	1.76 GB	9.556	98.6 %	默认：近无损
lfm3-1.7b-base-q4tp.cmf	4 位	0.93 GB	9.984	87.4 %	内存最小
bf16 原模型（Transformers）	—	3.46 GB	9.556	—	—

NVIDIA A40（Vulkan）上 q8_2f 生成 166 token/s、4k 提示词 657 token/s，q4tp 为 228 与 622 token/s；Apple M4（Metal，cortiq 0.8.50）上 q4tp 生成 84 token/s。这些文件是 Liquid AI LFM3-1.7B-Base 的修改形式（由 cortiq 转换为 CMF 并量化），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。