inclusionAI released Ling-3.0-flash-UBTree speculative-decoding draft model on Hugging Face
inclusionAI released Ling-3.0-flash-UBTree, a 482M-parameter speculative-decoding draft model for Ling-3.0-flash, on Hugging Face under the MIT License. The checkpoint is not standalone and requires the target model; SGLang integration is not yet publicly available.
China context
- Original name
- 蚂蚁集团
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders can download the weights under MIT, but cannot yet use the model for inference acceleration because the SGLang integration is not released. They should monitor the repository for serving instructions.
- For investors
- The release signals continued investment in inference optimization by Ant Group's AI arm, but the lack of benchmarks and serving support means the commercial impact is not yet measurable.
inclusionAI published the Ling-3.0-flash-UBTree checkpoint on Hugging Face. It is a UBTree speculative-decoding draft model designed to accelerate inference for the Ling-3.0-flash target model. The draft uses a five-layer DFlash backbone and a learned bigram selector to propose multiple candidate continuations for verification. The model weights are released under the MIT License. The repository notes that SGLang integration is not yet released, and benchmark results are not yet reported.
The draft model has 482M parameters, BF16 weights, hidden size 2560, 32 attention heads and 32 KV heads. It uses a draft block size of 16 and a default verification tree budget of 64 non-root nodes plus one anchor. The model shares token embeddings and the language-model head with the target, but those shared weights are not included in the checkpoint. The UBTree method pairs a unigram proposer with a bigram selector and uses tree-native training with renormalized KL supervision.
Developers using Ling-3.0-flash cannot yet deploy this draft model for faster inference because the SGLang integration is not publicly available; they must wait for installation instructions and serving recipes. The MIT license permits commercial use and modification once the serving tooling is released.
If the SGLang integration is released and benchmarks show meaningful speedup, this draft model could reduce inference costs for deployments of Ling-3.0-flash. The MIT license removes licensing friction for commercial adoption.
The next verifiable signal is the release of SGLang integration and benchmark results for acceptance length and speedup on datasets such as GSM8K, MATH-500, and HumanEval. Until then, the practical speedup of this draft model remains unverified.