infosave / cmf

CMF — Cortiq Model Format

One file. No Python, no torch, no CUDA install, no C++ toolchain.

A .cmf carries the quantized weights, the tokenizer and the chat template together, checks its own integrity, and memory-maps straight off disk. The runtime is a small Rust core with no ML framework under it, running on CPU everywhere and on GPU through wgpu — Vulkan, DX12, Metal — out of the box.

This repository is the format's front page. It holds the specification and the index of published models; it carries no weights of its own.

• Specification — SPEC.md , the normative document
• Source — github.com/infosave2007/cmf (Apache-2.0)
• Runtime — cortiq-cli on crates.io , and prebuilt binaries for six targets on the releases page

Run one

cargo install cortiq-cli                    # or take a prebuilt binary
hf download infosave/Nanbeige4.2-3Bcmf nanbeige42-3b-q4t.cmf --local-dir .
cortiq verify nanbeige42-3b-q4t.cmf         # per-tensor hashes
cortiq run nanbeige42-3b-q4t.cmf --prompt "Write a haiku about memory-mapped files."

Converting a checkpoint is one command and no Python:

cortiq convert --model Qwen/Qwen3-0.6B --quant q8 --output qwen.cmf

What is actually in the file

A CMF file is an envelope, not a tarball: readers navigate through the header and never guess at offsets. Unknown header fields are ignored, so the format grows additively — a breaking change costs a feature bit or a version bump, never a silent reinterpretation. A file written today stays readable, and cortiq verify is the contract that says so.

Beside the weights it carries the things that are usually scattered across a repository or lost entirely: the tokenizer and chat template, per-tensor hashes, an optional Ed25519 signature, per-task masks that select an active subset of the shared weights, and a swarm of skills sharing one backbone — several specialists in one file, chosen at load time rather than downloaded separately.

Two things it does that other single-file formats do not:

Attention that stops growing with the context. Converting with --o1 replaces a layer's softmax attention with a streaming operator holding a fixed-size state — a few exact anchor keys, an exact recent window, and a landmark sketch of everything older, under one shared denominator. The weights are byte-identical; the flag only records a hint in the header.

Generative models in the same container. The format is not LLM-only. cortiq imagine renders images and cortiq animate renders video with synchronized stereo audio, from .cmf files packed the same way, on the same runtime.

Published models

Every one of them runs from the same binary: cortiq run for the text models, cortiq imagine for Lumina, cortiq animate for MiniMax-H3.

Where it is honest

The format is settled at v2 and evolves additively. The crate APIs may still move before 1.0. Quality claims in the model cards are measured held-out numbers with the corpus named, not estimates — where a number was not measured, the card says so rather than guessing, and the container refuses to record a quality field without a measurement.