CPU3DAI tools for 3D, video and audio

Tencent releases draft model DSpine for Qwen3.8-27B

The model tencent/Qwen3.8-27B-DSpine-b16 with the text-generation task has been published on HuggingFace. It is a five-layer draft model for Qwen3.8-27B that generates draft tokens in parallel: each layer passes predicted features to neighboring positions, carrying causal dependencies through the depth of the network. The block size is 16, the sliding window is 8,192 tokens.

What it means

The model contains 2B parameters, the largest weights file is model.safetensors at 3.8 GB. The license is apache-2.0, the weights are published in safetensors format. There are no quantized builds on the platform yet. To run it, the author specifies SGLang 0.5.22 in BF16 with FlashAttention 3, as well as dynamic verification with a reference to DSpark and ReplaySSM enabled. This is not a standalone text model but a draft for accelerating inference of Qwen3.8-27B. In benchmarks, DSpine shows a larger accepted block length than the official DFlash2 — for example, 7.20 versus 5.69 on GSM8K with greedy sampling. For local running you will need to keep both the main model and this draft in memory, so you should estimate the combined size of both. More on tools for local running is in the CPU3D news feed.
DSpine: draft model for Qwen3.8-27B Qwen3.8-27B main model text-generation 27B parameters DSpine-b16 draft model 2B parameters 5 layers, block 16 Verification dynamic DSpark + ReplaySSM SGLang 0.5.22 BF16 request draft tokens Parallel draft-token generation each layer passes features to neighboring positions causal dependencies across network depth Accepted block length on GSM8K DSpine: 7,20 DFlash2: 5,69 Both models must be in memory: total weight size
How the method works. The diagram was drawn for this news note.

See also