speculative decoding

Tencent has published Qwen3.8-27B-DSpine-b16, a five-layer draft model with a 16-token block size and an 8,192-token sliding window, under Apache 2.0. According to its model card, it beats DFlash2 on acceptance length across five general benchmarks and on end-to-end agent throughput in SGLang.

KRAFTON AI has released Raon-1 mini preview, a Mixture-of-Experts language model trained from scratch with 28.8B total and 4.3B active parameters. It is Apache 2.0 licensed, supports 262,144 tokens of context, and ships with a speculative decoding draft model that KRAFTON says reaches up to 2.18× higher output throughput on a single NVIDIA B300.

MiniMax has published MiniMax-M3.1-Flash on Hugging Face: a native multimodal mixture-of-experts model with about 428B total parameters, 23B activated per token, and a context window of up to 1M tokens. It builds on MiniMax-M3 with sparse attention, quantization-aware training and DSpark speculative decoding.

Liquid AI has released LFM2.5-VL-3B-DSpark, a 280M-parameter speculative decoding drafter for its LFM2.5-VL-3B vision-language model. The company claims decode speedups up to 3.13x on Apple silicon and 2.66x on H100 GPUs, with day-one support for llama.cpp, MLX-VLM, and SGLang.

Liquid AI released DSpark draft model checkpoints for three LFM2.5 models, enabling speculative decoding that speeds up inference by up to 3.18x on H100 GPUs and 2.87x on-device, with day-one support for llama.cpp and SGLang.