Ncm

Ncm is an open-source text generation framework that uses a two-stage hierarchical model for parallel processing.

Next-Concept Modeling is a two-stage framework designed for parallel text generation. It utilizes a multiscale residual VQ tokenizer to represent a sequence as a hierarchy of discrete codes, and a transformer predicts that hierarchy from coarse to fine. Positions within each level are generated in parallel, meaning generation requires only one forward evaluation per level. The framework supports various tasks including summarization, machine translation, Sudoku puzzles, and language modeling on datasets like XSum, WMT14, and TinyStories. Released models and tokenized datasets are hosted on Hugging Face under the MIT license.

Ncm at a glance

Best for: Software engineer, Researcher
Runs on: Python, CLI, open source
Pricing: Free
Released: October 2026
Category: AI research tool
Website: github.com/hazyresearch/ncm

Key features

Coarse-to-fine text generation framework
Multiscale residual VQ tokenizer
Parallel generation of sequence positions
Per-GPU generation shard execution
Evaluation scripts for ROUGE, BLEU, and perplexity
Support for classifier-free guidance

Use it for

Text summarization on XSum
Machine translation on WMT14 De-En
Solving Sudoku puzzles
Language modeling on TinyStories and OpenWebText

Alternatives to Ncm

Text Generation WebUI — Run local LLMs with multiple backends, multimodal input, and tool calling 100% offline.
Overview — Learn language models by running hands-on experiments without writing scripts
NanoJev — It runs parallel decisions and outputs probability distributions without output-token decoding.
MiniCPM OCR — Efficient document parsing with correctness assessment and high throughput.
Anime Text Dataset — Provides structured anime data with field-level provenance and question-answering pairs.
ncnn — No runtime dependencies and supports CPU and Vulkan GPU backends.
NuMind — Achieves high accuracy in document extraction and outperforms generalist models.
LLM Wiki Research — Compiles parallel agent research into source-backed markdown knowledge bases.
Finetuning With Sampling — Official implementation of the Finetuning with Sampling research paper from Harvard.
Pagesift — Extracts web data with high memory efficiency and speed for AI agents.
MTEB — Evaluates embedding models across over 1000 tasks and 1000 languages.
UltraRAG — Build complex RAG workflows using modular MCP components and YAML configuration.

All Ncm alternatives