Cola2: novidades, hardware necessário, como rodar grátis

Cola2 blends continuous semantic compression with discrete token generation. According to the model card, it reaches a 60.10 % average on eight tasks after training on just 346 B source tokens, far fewer than the trillions used by many larger models.

Cola2 is a base language model released by ByteDance-Seed. It has 8.8 billion parameters and follows a continuous-latent architecture. The model uses a dynamic-compression variational auto-encoder to turn spans of text into compact continuous representations. Those latents are then processed by a block-causal Diffusion Transformer, and finally a span-causal token decoder turns the latents back into ordinary tokens. The repository provides the weights, tokenizer, and generation configuration, all ready to run with Hugging Face Transformers using remote code. The model is not instruction-tuned, so it behaves like a classic text generator rather than a chat assistant.

What sets Cola2 apart is its co-adaptive hierarchy. In Stage I the VAE learns to compress text into continuous latent spans while the token decoder learns to reconstruct them. Stage II adds a roughly 4 billion parameter Diffusion Transformer that learns the latent distribution through flow matching. During inference, the model proposes a latent block, decodes it into tokens, and then re-encodes the newly generated text under the same span layout. This feedback loop ensures the continuous and discrete representations stay synchronized, a design not found in standard autoregressive models.

The model card reports results on eight evaluation tasks. Cola2 achieves an aggregate average of 60.10 percent, which is lower than the 66.25 percent of Llama-3-8B but higher on specific benchmarks. It scores 41.75 percent on HellaSwag and 80.42 percent on StoryCloze, both better than Llama-3-8B’s 39.21 and 79.49 respectively. The checkpoint was trained on 346.03 billion source tokens, a fraction of the 15 trillion tokens used by Llama-3-8B, highlighting the efficiency of its compression approach.

Cola2 shows clear strengths in tasks that require narrative understanding and commonsense reasoning, such as StoryCloze and HellaSwag. However, its performance drops on knowledge-heavy benchmarks like MMLU (51.32 percent) and TruthfulQA (54.10 percent), where larger models still lead. Because latent generation starts from Gaussian noise, the model’s outputs are stochastic even when greedy token selection is used. Users can seed the random generator to obtain reproducible runs, but the inherent randomness remains a characteristic limitation.

Cola2 is a good fit for anyone interested in experimenting with continuous latent representations in language models. Researchers can study the co-adaptive VAE and diffusion components without building their own infrastructure. Hobbyists with a modest GPU, an RTX 4060 is enough to hold the model in memory, can run it locally and explore its unique generation style. Since the model is a base, not instruction-tuned, it suits tasks that need raw text generation rather than conversational agents.

O que você precisa para rodá-lo em casa? Com 4 bits, ele ocupa cerca de 8 gigabytes de memória. A menor máquina doméstica que o comporta por inteiro é RTX 4060. Em RTX 3060 12 GB com 32 gigabytes de RAM, ele escreve cerca de 48 tokens por segundo. Em RTX 4090 com 64 gigabytes de RAM, ele escreve cerca de 123 tokens por segundo.

There is no free API for it yet. The weights are open: you can download them from Hugging Face, under the license written in its model card. Arena has not ranked it yet: we add the score as soon as the votes arrive.

Compare Cola2 with every free model, and check what your own GPU can run, on FreeLLMArena.