# Cola2 (slim)

> Cola2 is a research checkpoint from ByteDance Seed, a base language model of 8.8B parameters that generates text through continuous latents with a diffusion transformer and a token decoder. Not a chat model, and the weights carry no stated licence.

- Full: https://www.anchorterminal.com/tools/cola2.md (~8,700 tokens) · this version ~1,630 tokens · JSON https://www.anchorterminal.com/tools/cola2.json · canonical https://www.anchorterminal.com/tools/cola2
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/cola2#pricing · reviews https://www.anchorterminal.com/tools/cola2#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**F · 33.9/100 · rank #160 of 160 open-weight models · #113 in Open-weight models · not agent-ready · confidence low**


Assessment: A research release, and an honest one: ByteDance Seed says it's a base model with an unconverged latent stage, scoring below Llama-3-8B on its own table. It's here for people studying continuous-latent generation. Agents can't use it as a model: 32-token completions, no chat or tools, the repository's own code to load it, and no licence on the weights.

## Facts

- Kind: Open-weight model · vendor: ByteDance (Seed) · category: Open-weight models · legal entity: ByteDance, named in the GitHub repository's security policy. The legal entity wasn't looked up · provenance 58/100
- Packages: huggingface `ByteDance-Seed/Cola2`
- Auth: None · pricing: Free · x402: no · licence: None stated for the weights (the code is Apache 2.0)
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/ByteDance-Seed/Cola2, read from the model card, the files and their metadata, the config, the tokenizer config, the commit history, and the LICENSE, SECURITY.md and docs/model_card.md of the GitHub repository, on 10 November 2026. ByteDance's company facts come from our BytePlus listings. The paper, the repository's Python files and the other docs weren't read. Nothing was downloaded or run
- Model: 8,768,019,912 parameters in the files. Three parts, per the GitHub model card: a dynamic-compression VAE that maps spans of 1 to 8 tokens to 64-dimensional latents, a block-causal latent diffusion transformer of about 4B parameters and 34 layers, and a 36-layer span-causal token decoder shaped like Qwen3-4B and initialised from Qwen3-4B-Base
- Files: Eight safetensors shards in F32 (35.07 GB), config and tokenizer files, three Python modules (configuration_cola2.py, modeling_cola2.py, modeling_layers.py), eight evaluation data files and SHA256SUMS. No pickle files. Hugging Face lists a SHA-256 for the large files
- Licence: None stated for the weights. The card's licence section says the repository's code is Apache 2.0 and that model weights, tokenizer assets and benchmark data remain subject to their respective terms, and points to a NOTICE file. The metadata has no licence field
- Runtimes: Transformers 4.51 to 4.x with trust_remote_code, per the card, which rules out version 5. Released vLLM 0.33.0, SGLang 0.5.23, Transformers 5.22.2 and llama.cpp b12242 have no cola2 code
- Context: Not stated. config.json gives 32,768 positions. The GitHub model card says practical limits depend on memory, and that its benchmark truncates prompts to 2,048 tokens and generates at most 32
- Hardware: No memory figure. The card recommends CUDA with bfloat16. The files are 35.07 GB in F32
- Training data: OLMo 2 pretraining text retokenised with the Qwen3-4B tokenizer, 346.03B source tokens over three stages (52.43B, 125.83B and 167.77B), per the GitHub model card, which says the corpus and training code aren't released. No cut-off
- Recommended settings: The released preset: an 8-step Euler schedule for the diffusion transformer, guidance scale 10, decoder guidance -0.9, greedy token selection and 32 new tokens. Generation starts from noise, so pass a seeded generator to compare runs

- Scores: Capability pending, Reach 25, Risk 40, Agent interface 20, Documentation and upkeep 58 · total over the 7 assessed categories
- Why:  · Reach, Memory 15, the only build is the F32 safetensors set, 35.07 GB, and 35.07 GB × 1.2 = 42.1 GB, within 64 GB. · Risk, File safety 15, safetensors only, no .bin, .pt or .pth files (25). · Agent interface, Chat template 20, A chat template in tokenizer_config.json (4,116 characters), which came with the Qwen3 tokenizer. · Documentation and upkeep, Model card 8, Intended use and limitations (the out-of-scope list, and the unconverged latent model) and the architecture are on the card an…
- Sources: 13, open questions: 7, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/cola2.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: https://github.com/ByteDance-Seed/Cola2; more in the full twin

## Before you call it

1. Don't plan to use this as your model. It completes short prompts and wasn't tuned to follow instructions
2. Don't use it commercially on the strength of this listing. The weights have no stated licence
3. If you study it, read modeling_cola2.py first: trust_remote_code runs it
4. Pin Transformers below 5, as the card requires
5. Pass a seeded torch.Generator to generate(). Output varies between runs even with greedy decoding
6. Check the download with `sha256sum -c SHA256SUMS`

## Connect

```bash
pip install "torch>=2.2" "transformers>=4.51,<5" "safetensors>=0.4" accelerate
# then AutoModel.from_pretrained("ByteDance-Seed/Cola2", trust_remote_code=True, torch_dtype=torch.bfloat16)
# trust_remote_code runs the three Python files in the repository. Read them first
```

```bash
# No server. The model runs in process through Transformers, and the GitHub repository has a
# command-line chat script (scripts/run_chat.sh) and a benchmark runner
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 75.4 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 73.6 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 12B IT | BB | 72.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)