# Cola2

> Cola2 is a continuous-latent language model that co-adapts semantic compression and discrete text generation. It combines a dynamic-compression VAE, a block-causal Diffusion Transformer (DiT), and a span-causal token decoder with latent feedback.

- **Organisation:** ByteDance-Seed
- **Category:** Chat & Text Generation, Reasoning Models
- **Parameters:** 8.8B
- **Context length:** 32,768 tokens
- **Licence:** Unknown licence (Check licence)
- **Released:** 2026-11-06
- **Hugging Face:** 1.4K downloads (30 days), 15 likes

## Estimated memory

| Quant | ≈ GB |
|---|---|
| Q2_K | 4.39 |
| Q3_K_M | 5.76 |
| Q4_K_M | 6.77 |
| Q5_K_M | 7.67 |
| Q6_K | 8.63 |
| Q8_0 | 10.68 |
| F16 | 18.72 |

## Runs with

- [Transformers](https://github.com/huggingface/transformers)

## About

Cola2 is a continuous-latent language model that co-adapts semantic compression and discrete text generation. It combines a dynamic-compression VAE, a block-causal Diffusion Transformer (DiT), and a span-causal token decoder with latent feedback.

GitHub · 中文说明 · Paper · Project page

The accompanying paper is titled “Cola2: Co-Adaptive Hierarchical Language Modeling with Continuous Semantic Compression and Balanced Discrete Generation.” Cola2 dynamically groups text into semantic spans, represents each span with a compact continuous latent, generates latent blocks with a DiT, and realizes those latents as discrete tokens.

The released checkpoint was trained on 346.03B source tokens and reaches a reported 60.10 average across eight evaluation tasks. These are reported paper results; individual runs may differ because latent generation is stochastic.

This repository contains:

• eight sharded safetensors weight files;
• tokenizer and generation configuration files;
• Hugging Face Transformers remote-code modules;
• normalized JSONL data for eight evaluation tasks; and
• SHA256SUMS for release-file verification.

The inference implementation, command-line tools, examples, benchmark runner, and documentation are maintained in the Cola2 GitHub repository. Training code is not included in this release.

Click the preview to watch the full video · Open the interactive project page

## Alternatives

- [ornith-ai/Ornith-1.0-9B](https://selfhost.directory/models/ornith-ai/ornith-1.0-9b) (9B, MIT, ≈6.84 GB at Q4): Aloha! 🌺 Today, we are releasing Ornith-1.0, a self-improving family of open-source models for agentic coding.
- [ornith-ai/Ornith-1.5-9B](https://selfhost.directory/models/ornith-ai/ornith-1.5-9b) (9.7B, MIT, ≈7.22 GB at Q4): Chirp Chirp! 🐦 We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement.
- [Qwen/Qwen3-8B](https://selfhost.directory/models/qwen/qwen3-8b) (8.2B, Apache 2.0, ≈6.48 GB at Q4): Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models.
- [ibm-granite/granite-4.2-8b](https://selfhost.directory/models/ibm-granite/granite-4.2-8b) (8.8B, Apache 2.0, ≈6.96 GB at Q4): Granite is a family of open-source large language models developed by IBM, designed for enterprise and research use. Granite models are built to be versatile, safe, and efficient — covering a range of sizes and capabilities from compact edge-deployable models to large-scale reasoning systems.
- [ibm-granite/granite-4.1-8b](https://selfhost.directory/models/ibm-granite/granite-4.1-8b) (8.8B, Apache 2.0, ≈6.96 GB at Q4): Model Summary: Granite-4.1-8B is a 8B parameter long-context instruct model finetuned from Granite-4.1-8B-Base using a combination of open source instruction datasets with permissive license and internally collected synthetic datasets.
- [IFM/K2-Horizon-7B](https://selfhost.directory/models/ifm/k2-horizon-7b) (9B, Apache 2.0, ≈6.96 GB at Q4): K2-Horizon-7B is the medium dense member of the K2-Horizon family: a 7B-core decoder-only model with a 512K context window.

More: https://selfhost.directory/models/bytedance-seed/cola2/alternatives

## Runs well on

- [GeForce RTX 4060](https://selfhost.directory/models/bytedance-seed/cola2/on/rtx-4060): Q4_K_M, est. 28 tok/s
- [Apple M4 (16 GB)](https://selfhost.directory/models/bytedance-seed/cola2/on/apple-m4-16gb): Q6_K, est. 9.1 tok/s
- [Radeon RX 6700 XT](https://selfhost.directory/models/bytedance-seed/cola2/on/rx-6700-xt): Q8_0, est. 23 tok/s
- [GeForce RTX 3060 12GB](https://selfhost.directory/models/bytedance-seed/cola2/on/rtx-3060): Q8_0, est. 21 tok/s
- [Arc B580](https://selfhost.directory/models/bytedance-seed/cola2/on/arc-b580): Q8_0, est. 27 tok/s
- [Tesla T4](https://selfhost.directory/models/bytedance-seed/cola2/on/tesla-t4): Q8_0, est. 19 tok/s

## Similar size

- [meta-llama/Llama-3.1-8B-Instruct](https://selfhost.directory/models/meta-llama/llama-3.1-8b-instruct) (8B)
- [Qwen/Qwen3-8B](https://selfhost.directory/models/qwen/qwen3-8b) (8.2B)
- [Qwen/Qwen2.5-7B-Instruct](https://selfhost.directory/models/qwen/qwen2.5-7b-instruct) (7.6B)
- [meta-llama/Meta-Llama-3-8B-Instruct](https://selfhost.directory/models/meta-llama/meta-llama-3-8b-instruct) (8B)
- [xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic](https://selfhost.directory/models/xtools-at/deepseek-v4-pro-qwen3.5-9b-uncensored-heretic) (9.4B)
- [google/gemma-2-9b-it](https://selfhost.directory/models/google/gemma-2-9b-it) (9.2B)

## FAQ

**How much VRAM does Cola2 need?**
About 6.8 GB at Q4_K_M with an 8K context. Q8_0 needs about 11 GB and full precision about 19 GB. Longer contexts need more memory for the KV cache.

**Can Cola2 run on the RTX 4060?**
Yes. It fits at Q4_K_M (about 6.8 GB), at an estimated 28 tokens/s. It fits comfortably on the Apple M4 (16 GB), Radeon RX 6700 XT, RTX 3060 12GB.

**Can I use Cola2 commercially?**
Unclear. It is released under Unknown licence. No licence stated: assume all rights reserved. Check the licence on the model card before using it in a product.

**What is the context length of Cola2?**
32,768 tokens (32K). The full window needs far more memory than an 8K context.

**How do I run Cola2 locally?**
It runs with Transformers. Each has copy-paste examples in its docs.

## Links

- huggingface: https://huggingface.co/ByteDance-Seed/Cola2
- homepage: https://hongcanguo.github.io/Cola2/
- github: https://github.com/ByteDance-Seed/Cola2
- paper: https://arxiv.org/abs/2610.31305

---
Source: https://selfhost.directory/models/bytedance-seed/cola2