Published: 21 August 2026
Liquid AI released LFM2.5-DSpark on 20 August 2026, three speculative decoding draft models that make its LFM2.5 family up to 3.18x faster on an H100 without changing a single output token. The drafters add roughly 300M parameters each and attach to LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-8B-A1B. Weights are on Hugging Face today.
What is LFM2.5-DSpark and how does it work?
Speculative decoding uses a small model to guess the next tokens and a larger model to check them in a single pass. LFM2.5-DSpark applies that to Liquid AI’s own family: a five-layer drafter of roughly 300M parameters proposes a block of nine candidate tokens, and the target model verifies the whole block at once.
Under greedy decoding the emitted sequence is identical to the target model running alone, so benchmark accuracy does not move. The cost is memory, and the LFM2.5-2.6B drafter repository is 655 MB in BF16. Liquid AI trained the drafters exclusively on AMD hardware, and llama.cpp and SGLang both shipped support on day one.
LFM2.5-DSpark benchmarks: 3.18x on an H100, 2.87x on a MacBook
- On a single H100 in BF16 through SGLang, LFM2.5-2.6B averages 2.67x, moving from 323 to 864 tokens per second. LFM2.5-8B-A1B peaks at 3.18x on MATH500, from 428 to 1362 tokens per second.
- On an M4 Max MacBook Pro through llama.cpp, LFM2.5-1.2B-Instruct hits 2.87x on HumanEval, from 136 to 389 tokens per second.
Speedup follows the acceptance rate, so it swings with the workload. LFM2.5-8B-A1B accepts 8.27 of 10 drafted tokens on MATH500 but only 4.02 on GSM8K, dropping the same model to 1.29x on the same GPU. Every figure is Liquid AI’s own, measured at batch size 1 and temperature 0 across MATH500, GSM8K, HumanEval, MBPP and MT-Bench.
How LFM2.5-DSpark compares to EAGLE-3 and DFlash
DSpark is not Liquid AI’s invention. It comes from a July 2026 paper and builds on EAGLE-3, the method most speculative decoding stacks use today, and on DFlash, which drafts a whole block in one forward pass. DSpark adds a Markov-chain head that makes each drafted token depend on the one before it, raising acceptance at later block positions.
For a team already on LFM2.5 the alternative is simply running the target alone, and DSpark wins by roughly 2x for 655 MB of extra memory. The wider point is that Liquid AI publishes and maintains the drafter as part of the model family. Speculative decoding normally falls to whoever deploys the model, which is why most on-device stacks never switch it on.
Licence, availability and the mixture-of-experts weak spot
Apple silicon is the one place DSpark barely helps. LFM2.5-8B-A1B gains 1.18x on average on the M4 Max against 2.54x on the H100. Liquid AI blames the current mixture-of-experts implementation in llama.cpp’s Metal backend, plus the fact that verifying nine tokens activates more experts than a single decode step, and publishes the weak numbers anyway.
The blog describes the weights as open with no restrictions, but the LFM Open License v1.0 is narrower: free commercial use applies only while your organisation stays under $10M in annual revenue, and above that you need a commercial licence from Liquid AI. Checkpoints ship as Safetensors and GGUF, and no hosted provider serves them, so this is a self-hosting release.
What this means
Worth testing now if you already run LFM2.5 on your own hardware, and safe to ignore otherwise. The drafters bind to specific targets, so there is nothing here for a team on Qwen3.8, Gemma 4 or a hosted API. It earns its keep in on-device agents: a 57% latency cut on multi-tool function calling for LFM2.5-2.6B changes how an app feels.
The weak mixture-of-experts result deserves credit rather than criticism, because publishing a 1.18x figure next to a 3.18x headline tells you where the engineering still is. Our reservation is the licence. A $10M revenue ceiling on free commercial use is awkward for exactly the mid-sized companies with on-premise requirements that Liquid AI is courting, so read the terms before you build on it.
For more information you can visit the official LiquidAI announcement of the DSpark draft models.
