Liquid AI’s draft models cut LFM2.5 decoding by 3.18x

Liquid AI has released DSpark draft checkpoints for three LFM2.5 models, promising up to 3.18x faster decoding without changing the generated output. Marktechpost reports that the release targets LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-8B-A1B.

DSpark uses speculative decoding: a roughly 300 million-parameter drafter proposes a block of nine tokens, then the larger target model verifies the entire block in one forward pass. Under greedy decoding, the accepted sequence is identical to the target model running alone, so benchmark accuracy remains unchanged. The trade-off is additional memory, including 655 MB for the LFM2.5-2.6B drafter in BF16.

Liquid AI measured the models at batch size 1 and temperature 0 using SGLang on an H100 and llama.cpp with Metal on an M4 Max MacBook Pro. Results varied substantially by model and workload:

• LFM2.5-1.2B-Instruct: 2.10x faster on average on H100, reaching 2.56x; 2.54x faster on average on M4 Max, reaching 2.87x.
• LFM2.5-2.6B: 2.67x faster on average on H100, reaching 3.06x; 2.27x faster on average on M4 Max, reaching 2.63x.
• LFM2.5-8B-A1B: 2.54x faster on average on H100, reaching 3.18x; just 1.18x faster on average on M4 Max, reaching 1.44x.

The speedup depends on how many proposed tokens the target accepts. For LFM2.5-8B-A1B, acceptance reached 8.27 of 10 tokens per step on MATH500 but fell to 4.02 on GSM8K, reducing the gain on the same H100 from 3.18x to 1.29x. On MT-Bench, the 1.2B model’s acceptance rate dropped to 3.90 and its H100 speedup fell to 1.66x.

Recommended reading

OpenAI tightens AI security after Hugging Face breach

Sergey Kuznetsov • Aug 19, 2026 • 2 min read

The weak point is the 8B-A1B mixture-of-experts model on Apple silicon. Liquid AI attributes the modest M4 Max improvement to llama.cpp’s current Metal MoE implementation: verifying multiple tokens activates more experts and creates more weight traffic than a single decode step.

Self-hosting is required

The checkpoints are available in Safetensors and GGUF formats, with support in llama.cpp and SGLang. They are not currently served by a hosted inference provider on Hugging Face, so deployment requires a compatible local build and the DSpark-enabled LFM2 target.

For agent workloads, Liquid AI reports a more practical benefit: DSpark reduced latency by 57% on average for multi-tool function-calling scenarios using LFM2.5-2.6B. That use case repeatedly pays decoding costs while an agent plans, calls tools and revises its response.

The models use the LFM Open License v1.0, which permits free commercial use for entities with less than $10 million in annual revenue. Larger companies must obtain a commercial license from Liquid AI before using the checkpoints commercially.