Liquid AI introduces dspark draft models for faster decoding without compromising output accuracy

Liquid AI has launched draft models for its LFM2.5 series, including LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. These models incorporate speculative decoding to significantly enhance performance. The drafters, each with around 300 million parameters, propose a block of nine candidate tokens, which the target model verifies in a single forward pass. This approach results in decoding speeds up to 3.18 times faster on an H100 GPU and up to 2.87 times faster on an M4 Max MacBook Pro. Output remains unchanged, ensuring benchmark accuracy is preserved. Both llama.cpp and SGLang support the models, though they require a build with DSpark support for LFM2 targets. The speculative decoding mechanism uses a small model to generate tokens that a larger model then validates. The drafters are built with a backbone of five full-attention layers, hidden_size of 2048, intermediate_size of 6144, and GQA at 32 heads over 8 KV heads. The drafters do not include vocabulary weights, with embeddings and LM heads tied from the target model at load time. The 2.6B drafter repository is 655 MB in BF16, representing the real memory cost added. DSpark combines three components: a DFlash-style parallel backbone that produces hidden states for all draft tokens in one forward pass, a lightweight sequential head that restores inter-token dependencies, and a confidence-scheduled verifier that predicts token survival probabilities and prunes low-confidence suffixes. Liquid AI reports throughput on 1xH100 in BF16 via SGLang and on an M4 Max MacBook Pro via llama.cpp with Metal and FP16 GGUF weights. Both use a block size of 9, batch size of 1, and temperature of 0 across multiple benchmarks including MATH500, HumanEval, MBPP, GSM8K, and MT-Bench. Speedup is closely tied to the acceptance rate, which reflects output predictability. For LFM2.5-8B-A1B, the model accepts 8.27 out of 10 tokens per step on MATH500 but only 4.02 on GSM8K, leading to a speedup drop from 3.18x to 1.29x on the same GPU. On the 1.2B model, MT-Bench acceptance drops to 3.90, reducing the H100 gain to 1.66x. On Apple silicon, the MoE result shows a clear limitation, with LFM2.5-8B-A1B achieving only 1.18x speedup on the M4 Max. Liquid AI attributes this to the current MoE implementation in llama.cpp’s Metal backend and the increased weight traffic from verifying multiple tokens. The draft model proposes a block of nine tokens after the anchor, and the target model checks all ten in a single forward pass. Matching tokens are retained, while the first mismatch is replaced by the target’s own token, with the tail discarded. Both lanes emit identical greedy sequences, with playback running at 1/20 speed to visualize the gap. The drafter does not include vocabulary weights, with embeddings and LM heads tied from the target model at load time.